Datasets:

Translation-EnKo

/

arxiv-paper

like 0

**Yi: Open Foundation Models by 01.AI** **01.AI** **Code:** [https://github.com/01-ai/Yi](https://github.com/01-ai/Yi) **Model:** [https://huggingface.co/01-ai](https://huggingface.co/01-ai) **Abstract** We introduce the Yi model family, a series of language and multimodal models that demonstrate strong multi-dimensional capabilities. The Yi model family is based on 6B and 34B pretrained language models, then we extend them to chat models, 200K long context models, depth-upscaled models, and vision-language models. Our base models achieve strong performance on a wide range of benchmarks like MMLU, and our finetuned chat models deliver strong human preference rate on major evaluation platforms like AlpacaEval and Chatbot Arena. Building upon our scalable super-computing infrastructure and the classical transformer architecture, we attribute the performance of Yi models primarily to its data quality resulting from our data-engineering efforts. For pretraining, we construct 3.1 trillion tokens of English and Chinese corpora using a cascaded data deduplication and quality filtering pipeline. For finetuning, we polish a small scale (less than 10K) instruction dataset over multiple iterations such that every single instance has been verified directly by our machine learning engineers. For vision-language, we combine the chat language model with a vision transformer encoder and train the model to align visual representations to the semantic space of the language model. We further extend the context length to 200K through lightweight continual pretraining and demonstrate strong needle-in-a-haystack retrieval performance. We show that extending the depth of the pretrained checkpoint through continual pretraining further improves performance. We believe that given our current results, continuing to scale up model parameters using thoroughly optimized data will lead to even stronger frontier models. ###### Contents * 1 Introduction * 2 Pretraining * 2.1 Data Processing * 2.2 Tokenization * 2.3 Model Architecture * 3 Finetuning * 3.1 Data Preprocessing * 3.2 Training Method * 4 Infrastructure * 5 Safety * 6 Evaluations * 6.1 Base Model Performance * 6.1.1 Main Results * 6.1.2 Discussions * 6.1.3 In-Context Learning Study * 6.2 Chat Model Performance * 6.2.1 Automatic Evaluations * 6.2.2 Human Evaluations * 7 Capability Extension * 7.1 Long Context Modeling * 7.2 Vision-Language * 7.3 Depth Upscaling * 8 Final Discussions * A Author List and Contributions Introduction Recent breakthroughs in large language models have revolutionized the whole field of artificial intelligence and potentially radiate across the entire human society. Our vision for large language models is to make them the next generation computational platform and empower the whole community with significantly amplified intelligence. As a step towards this mission, we present the Yi model series, 6B and 34B language models pretrained from scratch on 3.1T highly-engineered large amount of data, and finetuned on a small but meticulously polished alignment data. Due to the data quality resulting from our substantial engineering efforts, which we will detail in the upcoming sections, Yi achieves near GPT-3.5 benchmark scores and human preferences. In designing the Yi model series, we are mostly concerned on the following dimensions regarding _model scale, data scale, and data quality_: (1). when choosing model scale, the desiderata is to have small enough model that is feasible for inference on consumer-grade hardware like the RTX 4090 where the bounding factor is its limited 24G memory, yet still large enough with complex reasoning and emergent abilities. This is why we found 34B gives a nice performance-cost balance; (2). since 34B is smaller than the conventional 70B used by Chinchilla [30] and LLaMA [77], we increase the pretrain data scale to 3.1T tokens to compensate for the decreased compute flops. This makes the model-data scale combination fall into the post Chinchilla optimal regime [64], i.e., we overtrain the model on more tokens (3T) than the compute optimal (around 1T). The benefit is from the inference side, as we achieve stronger performance with reduced serving cost: after int4 [81] quantization, one can serve the 34B chat model on 24G GPU memory with almost no performance drop; (3). our data engineering principle is to promote quality over quantity for both pretraining and finetuning. The pretraining data quality is guaranteed by a sophisticated data cleaning pipeline with cascaded filtering methods and intentionally increased deduplication strength; (4). for finetuning data we heavily emphasize quality by handcrafting less than 10K instructions over multiple iterations based on user feedback. This approach significantly deviates from the quantity-scaling styled instruction tuning works like FLAN [9] and UltraChat [19], but aligns more with the handcrafting styled works like LIMA [94]. Our pretraining data cleaning system features a sophisticated filtering pipeline based on language, heuristic textual features, perplexity, semantics, topic, and safety, as well as a cascaded deduplication process based on paragraph, MinHash, and exact matching. This thorough pipeline leads to a much higher removal ratio than existing pipelines like CCNet [80], RefinedWeb [56] and RedPajama [13], which we believe is key to the success of data engineering. The underlying principle is although pretraining requires data scaling, one would like to make sure the data used are of high quality, rather than training the model on large raw data, i.e., we prefer 3T tokens over sophisticated engineering over 10T tokens without extensive filtering. Regarding the model architecture, we use standard implementation of the Transformer architecture with Grouped-Query Attention (GQA) [1], SwiGLU [68] activation, and RoPE with an adjusted base frequency (RoPE ABF) [82]. This design choice is the standard approach rooted from the Transformer original paper [78], later modified by GPT-3 and Chinchilla [30], then followed by LLaMA [77], Baichuan [84], Qwen [3] and many related works. To approach GPT-3.5-matching human preferences, our finetuning dataset is curated from carefully selected multi-turn instruction-response pairs, annotated directly by our team of machine learning engineers then polished over multiple iterations of user feedback. As mentioned above, the size of our finetuning dataset is less than 10K, but improved over and over again across the model development timeline. Benefiting from the dataset's manageable size, we employed an extensive grid search to identify the optimal data composition, promote diversity, and discover effective hyperparameters. After 8-bit and 4-bit quantization, the final chat model can be deployed on consumer-grade GPUs nearly without performance degradation compared to the bf16 format. We further extend the Yi model capability from three dimensions: context scaling, vision-language adaptation, and depth-upscaling. To achive 200K context length, we continue pretrain the model on about 5B length-upsampled data, similar to the concurrent work in Fu et al. [22]. To adapt the model to vision-language tasks, we integrate a vision encoder and develop a multi-stage training method, following and improving the practice of Liu et al. [47]. We also study the effectiveness of depth-upscaling [38], i.e., making the model deeper by continual pretraining, and confirming its effectiveness to further improve model performance. from pretraining to finetuning to serving. To support pretraining, we develop cross-cloud elastic task scheduling, automatic failure recovery, and topology-aware resource allocation which collectively enable us to run tasks according to the real-time available GPU nodes cross clusters with limited switching overhead. To support finetuning, we build a hierarchical scheduling framework supporting different distributed backends for different models (e.g., Megatron [70] for the policy model and DeepSpeed [60] for the reward model). For efficient inference, we use 4-bit model and 8-bit KV cache quantization, combining with PagedAttention [41] and Dynamic Batching. Extensive experiments demonstrate that Yi-34B can match GPT-3.5 in both performance and efficiency. On most standard benchmarks like MMLU [27] (for the base model) and LMSys ELO Rating [93] (for the chat model), Yi-34B generally achieves scores on par with GPT-3.5. After model parameter and KV cache quantization, the inference cost is also controlled such that a wide range of the community can deploy the model on cost effective devices. We further report a detailed performance comparison between Yi and major LLMs on commonsense reasoning, college exams, math, coding, reading comprehension, and human preference win-rate on multiple evaluation benchmarks. Since its release, the Yi model series has benefited the community from the following perspectives: (1). it provides GPT-3.5-matching quality yet cost-effective models to researchers, and enables developers to build AI-native applications like language model based agents; (2). it empowers end users with locally runnable chatbots, which consequently helps protecting user data privacy; (3). it sheds light on the direction on further data and model scaling to achieve even stronger frontier models. for both research and commercial use. ## 2 Pretraining Our approach to pretraining is to train a standard dense transformer architecture on a heavily engineered large pretraining corpora, where our underlying assumption is that when trained on extensive data of high-enough quality, a standard architecture can exhibit advanced capability. This is to say, we may not need much architectural modification, although we have indeed conducted extensive preliminary architectural experiments. In the following subsections, we first detail our data engineering pipeline, then briefly discuss the model architecture. ### Data Processing The Yi data mixture is shown in Fig. 2. To produce a high-quality bilingual pretraining data, we meticulously designed a cascaded data-processing pipeline, as illustrated in Fig 1. This pipeline features a series of data-cleaning strategies targeting quality and diversity. We start with web documents from Common Crawl, use the CCNet pipeline [79] for language identification and perplexity scoring. Then we use a combination of filtering and deduplication process, as detailed below. Figure 1: Yi’s pretraining data cleaning pipeline. Heuristic Rule FiltersThis part of filter aims for removing text of low quality. We filter out text based on: (1). URL, domain, word blocklists and garbled text filters; (2). document length, the ratio of special symbols, and the ratio of short, consecutive, or incomplete lines; (3). repeated words, n-grams, or paragraphs [58]; The filtering thresholds are based on a statistical analysis of large document samples, as described in Nguyen et al. [52]. Furthermore, we identify and anonymize Personal Identifiable Information (PII), such as email addresses and phone numbers. Learned FiltersWe use learned filters to address nuanced cases that exceed the capabilities of standard heuristic rules. Notably, the Chinese content extracted from Common Crawl present unique challenges, particularly with a higher ratio of inappropriate content like pornography and gambling. Traditional heuristic-rule-based filters struggle to effectively identify and eliminate all harmful content. To enhance our filtering process, we have integrated a suite of learned scorers for filtering, namely the perplexity scorer, quality scorer, safety scorer, and document coherence scorer: (1). the _Perplexity Scorer_, utilizing the KenLM library as per CCNet [80], evaluates a vast array of web documents, discarding those with perplexity scores largely above average; (2). the _Quality Scorer_ is a classifier trained to recognize and favor pages similar to Wikipedia in quality and assign scores accordingly. Documents that fail to meet the quality standard are subsequently removed; (3). the _Document Coherence Scorer_ identifies low-quality web documents that consist of disparate sentences or paragraphs, thus being incoherence. Such documents are either segmented for further analysis or removed entirely. (4). the _Safety Scorer_ identifies and removes web documents containing toxic content, such as violence, pornography, and political propaganda. Cluster-based FiltersWe further use unsupervised semantic clustering to group web documents. This clustering process enables efficient identification and analysis of documents sharing similar semantic features. The clustered data are subsequently annotated with quality labels, providing essential references for the optimization of Yi's data mixture strategy. Documents identified as low-quality through automatic and manual verification are excluded from the dataset. DeduplicationAfter filtering, we implement a comprehensive deduplication pipeline following the procedure in Penedo et al. (2023) [56]. This pipeline integrates document-level MinHash deduplication and sub-document exact-match deduplication, effectively identifying and removing duplicate content within and across documents. We further categorize web documents into specific themes using a topic model predicting labels like as news, ads, and knowledge-based content. In the final pretraining dataset, we down-sample less helpful content, mostly advertisements, to ensure information density. The final composition of Yi's pretraining data is shown in Fig. 2. ### Tokenization We use byte-pair encoding (BPE) [69] implemented in the SentencePiece framework [40], to tokenize the pretraining data. The vocabulary size of Yi is set to 64,000 to balance computational efficiency and word comprehension. Specifically, we split numbers into individual digits to facilitate a better understanding of numeric data. We allow rare characters to fall back to the unicode-byte encoding to ensure fault tolerance. We employ the identity tokenizer to avoid transferring all punctuations to the half-width format. LLMs prioritizing English usually utilize dummy prefix (whitespace at the beginning of text) in their tokenizers to generalize the same words at different positions of sentences. We do not use this approach because the assumption does not always hold even in the English context, especially for sentences that begin with quotation marks, also it does not show positive effect in Chinese context. Figure 2: Yi’s pre-training data mixture. Overall our data consist of 3.1T high-quality tokens in Both English and Chinese, and come from various sources. Our major differences from existing known mixtures like LLaMA [76] and Falcon [56] are that we are bilingual, and of higher quality due to our more rigorous cleaning pipeline. ### Model Architecture Yi uses a modified version of the classical decoder-only Transformer architecture [78] where the code is based on LLaMA's [77] implementation. The main parameter setting is summarized in Table 1. The modifications from LLaMA to Yi are further summarized below: Attention MechanismLLaMA 2 uses Grouped-Query Attention(GQA) [1] only on its largest 70B model, and its 7B and 13B uses full attention. We incorporate GQA in both Yi-6B and Yi-34B. GQA splits query-heads into G groups, sharing a single key and value head within each group of query [1]. This approach offers substantial reductions of training and inference costs, compared to the original Multi-Head Attention (MHA) [16; 57; 67]. We do not observe performance degradation after applying GQA to our 6B smaller model. Activation FunctionWe use SwiGLU [68] as Yi's post-attention layer, reducing its activation size from \(4h\) to \(8/3h\) (\(h\) denotes hidden size) to be consistent with the normal post-attention layer. This adjustment also compensates for the reduction in parameter resulted from GQA, making the overall parameter count compatible of existing 7B and 34B models. Positional Embedding and Long ContextWe use Rotary Position Embedding (RoPE) [73] following the standard implementation. We adjust the base frequency (RoPE ABF), introduced in Xiong et al. [82], to support long context windows up to 200K where the base model itself is trained on 4K context length. To adapt the base model to longer context, we continue pretrain the model on 10B tokens from our pretraining data mixture with slightly upsampled long sequences, mostly from book. We observe that only 1-2B tokens is enough for the model to converge to low loss on 4K-200K length, and a lightweight finetuning further induces near-perfect long-context retrieval performance. Based on this observation, we tend to view that the capability of modeling longer dependency than the pretrained length (4K) is a intrinsic capability (rather than an being injected by post-train). This is to say, the base model already has the capability to model longer than 4K dependency even the model is trained shorter, and the post-train / finetuning procedure simply release this capability. ## 3 Finetuning Our finetuning method significantly emphasizes data quality over quantity. Our approach does _not_ follow existing data-intensive approaches like FLAN [9] and UltraChat [19], which scales the SFT data to millions of entries but each of the entries may not been examined carefully because the scale is too large. In contrast, our method aligns with the LIMA [94] and DEITA [48] approach, which focus on data selection rather than scaling. With the scale being less than 10K, we are able to examine and optimize _every single data point_. Below we discuss our data construction and training details. ### Data Preprocessing Quality is All You NeedOur finetuning dataset consists of less than 10K multi-turn instruction-response dialog pairs, with each and every one of the entry constructed and polished over multiple iterations and from user feedback. We take this approach because in our preliminary experiments, we observe that compared to the open-source data of several hundred thousand entries, the results from a smaller, manually annotated dataset are superior. These observations align with those reported in Gemini Team et al. [23], Touvron et al. [77], Zhou et al. [94]. \begin{table} \begin{tabular}{c c c c c c c} \hline \hline Models & Hidden Size & Q-heads & KV-heads & Layers & Pretrain Seq. Len & Max LR \\ \hline 6B & 4096 & 32 & 4 & 32 & 4096 & \(3\times 10^{-4}\) \\ 34B & 7168 & 56 & 8 & 60 & 4096 & \(1.5\times 10^{-4}\) \\ \hline \hline \end{tabular} \end{table} Table 1: Model configs of Yi-6B and Yi-34B. LR stands for learning rate. We use the following techniques to improve prompt distribution selection, response formatting, and chain-of-thought formatting: (1). for prompt distribution selection, drawing inspiration from WizardLM[83], we develop compound instructions and progressively evolved them to increase their complexity. This approach has significantly reduced the size of SFT data in our experiments; (2). for response formatting, we generally use a default style extended from LIMA[94]. Overall, the responses are structured in an introduction-body-conclusion format where the body is usually a list of bullet point; (3). for CoT data formatting, we have use a "Step-Back" pattern, inspired by Zheng et al. [92], by performing abstraction to formulate higher-level solutions before delving into reasoning about the original, more concrete questions. We spend extra efforts on reducing hallucination and repetition: (1). to reduce hallucinations, we examine and ensure that the knowledge in the responses is not contained within the model, and eliminate responses that might lead to memorization; (2). to reduce repetition, we rewrite the repetitive turns of the responses that usually exist but may be overlooked in the finetuning data. Diversity and MixtureTo ensure the coverage of different capabilities, we have included a wide spectrum of open-source prompt, encompassing areas such as question answering, creative writing, dialogue, reasoning, mathematics, coding, safety, bilingual capabilities, and others. To obtain a fine-grained control of different directions of capabilities, inspired by InsTag[49], we develop a instruction tagging system. By designing a diversity-focused sampling algorithm, we carefully balanced the distribution of instructions across various tags. This approach ensures a diverse finetuning dataset, aiming to achieve enhanced cross-task robustness. To achieve the optimal data ratio for balancing different directions of the capability, we use an approximate grid search to determine our data mixture. Motivated by Dong et al. [20], this process involved experimenting with {1, 1/2, 1/4, 1/8, 1/16, 1/32, 1/64} proportions for each ability. The search process was guided by validation results and our in-house human evaluation sets. ChatML FormatBeyond the focus on data quality and diversity, our observations revealed that the format of the data substantially influences the model's ultimate performance. To this end, we implemented the ChatML-style format [53]. This structured approach empowers the model to differentiate among various information types, such as system configurations, user inputs, and assistant responses. ### Training Method We use next-word prediction loss for finetuning, and only compute loss on the responses, but not system and user instructions. We use AdamW optimizer with \(\beta_{1}\) set to 0.9, \(\beta_{2}\) set to 0.999, and \(\epsilon\) set to \(10^{-8}\). We use a sequence length of 4096, alongside a batch size of 64. We set training step to 300 with a constant \(1\times 10^{-5}\) learning rate, a weight decay of 0.1, gradient clipping with a maximum threshold of 1.0, and NEFTune [34] with a noise scale of 45 for Yi-34B-Chat and 5 for Yi-6B-Chat. ## 4 Infrastructure We build the infrastructure supporting the full-stack data processing, pretraining, finetuning, and serving. Our infrastructure features: (1). automated managing and monitoring the computing resource; (2). improved the training speed from optimized parallel strategies, kernel efficiency, and long-context support; (3). unified finetuning framework supporting heterogeneous distributed training backend, such as simultaneously using Megatron and DeepSpeed for multiple models in Direct Preference Optimization (DPO) [59]; (4). reducing the deployment cost by various LLM serving accelerations such as quantization, continuous batching, and paged attention. Below we explain these techniques one by one. Computing Resources ManagementTo efficient schedule large-scale language model development, particularly pretraining, which may take months on thousands of GPUs, we build a highly efficient multi-cloud task scheduling algorithm to manage pre-training, SFT, and RLHF tasks of different priorities. We also build a high-performance in-house training framework that allows us to automatically elastic scale the pre-train jobs to different node sizes based on the GPU availability. More importantly, all the training-related hyper-parameters will be scaled at the same time seamlessly. During the large language model training stage, a wide range of failures regularly occur, ranging from GPU crashes to communication fabric errors to loss spikes. We use the following strategies to address these reliability challenges: (1) we apply automated inspection, prediction, and labeling of nodes for different kind of software/hardware error categories. Nodes marked as tainted will be temporarily removed from the resource pool until the errors got cleared. (2) we implement a task queuing system with pre-checks and the capability for fast, automatic recovery in the event of failures during training tasks. (3) we develop of a user-friendly multi-task submission and management console, enabling developers to seamlessly manage and track their training tasks and hyper-parameters. Performance and Cost Efficiency_Memory_ and _communication_ restrictions are the two major technical challenges of large scale model training requiring integrated solutions beyond adding more GPUs. We use and improve upon the following techniques to tackle the memory and communication restrictions: (1) ZeRO-1 [60] to remove the memory consumption by partitioning optimizer states cross data-parallel processes; (2) tensor parallel combined with pipeline parallel [70] within each compute node to avoid inter-node communication bottleneck, and the 3D parallel strategy is well designed and optimized to avoid using activation checkpointing and minimize the pipeline bubbles; (3) kernel fusion techniques like flash attention[15][14] and JIT kernels to reduce redundant global memory access and consumption; (4) topology-aware resource allocation (ranking strategy) to minimize the communication across different layers of switches, which is the limitation of a typical fat-tree-topology. Finetuning FrameworkDifferent from pretraining, finetuning LLMs may require the orchestration of multiple models, as is the practice of DPO [59] and PPO [54]. In such training jobs, a typical process is to use reference/reward model to predict a batch of data (which also requires nontrivial time), then let the target model use this data to calculate loss and update parameters. To this end, we build a multi-model scheduling framework to support multiple backends for different LLMs in a single job. For example, when finetuning a language model with DPO, the intermediate results from the reference model can be cached and reused, improving the training speed and resource cost to be close to the supervised finetuning counterparts. Fast and Efficient InferenceWe primarily use quantization, dynamic batching, and Paged Attention for improving decoding speed and memory usage. We use quantization to decrease both the memory footprint and computation demand. By 4-bit model quantization [81] and 8-bit KV cache quantization [18], we are able to achieve significant GPU memory saving with near-zero performance degradation (e.g., less than \(1\%\) accuracy drop in MMLU/CMMLU benchmark). We use dynamic batching [86] to minimize the response time and improve batching efficiency. We use PagedAttention[41] to improve memory utilization and improve decoding. \begin{table} \begin{tabular}{l c c c c c c c c c c} \hline \hline & **Size** & **MMLU** & **BBH** & **C-Eval** & **CMMLU** & **Gaokao** & **CR** & **RC** & **Code** & **Math** \\ \hline **GPT-4** & - & **83.0** & **86.7** & 69.9 & 71.0 & 72.3 & **89.3** & - & **65.3** & **66.1** \\ **GPT-3.5** & - & 69.1 & 70.1 & 52.5 & 55.5 & 51.1 & 83.1 & - & 54.8 & 35.6 \\ \hline **Qwen** & 14B & 66.7 & 53.4 & 72.1 & 71.0 & 62.5 & 74.2 & 72.5 & 40.6 & 43.1 \\ \hline **Llama2** & 34B & 62.6 & 44.1 & - & - & - & 71.1 & 68.9 & 27.8 & 24.2 \\ & 70B & 69.7 & 64.9 & 50.1 & 53.3 & 23.3 & 72.7 & 72.3 & 38.4 & 35.2 \\ \hline **Baichuan-2** & 13B & 55.0 & 49.0 & 59.0 & 61.97 & 45.6 & 66.3 & 62.4 & 23.4 & 16.1 \\ **InternLM** & 20B & 62.1 & 52.5 & 58.8 & 59.0 & 45.5 & 78.3 & - & 34.8 & 30.26 \\ \hline **Skywork** & 13B & 62.1 & 41.7 & 60.6 & 61.8 & 68.1 & 72.4 & 61.4 & 64.9 & 18.1 \\ \hline **Falcon** & 180B & 70.4 & 54.0 & 57.8 & 58.0 & 59.0 & 74.4 & - & - & - \\ \hline \multirow{2}{*}{**Yi**} & 6B & 63.2 & 42.8 & 72.0 & 75.5 & 72.2 & 72.2 & 68.7 & 21.1 & 18.6 \\ & 34B & 76.3 & 54.3 & **81.4** & **83.7** & **82.8** & 80.7 & **76.5** & 32.1 & 40.8 \\ \hline \hline \end{tabular} \end{table} Table 2: Overall performance on grouped academic benchmarks compared to open-source base models. **CR** stands for Commonsense Reasoning. **RC** stands for Reading Comprehension. Long-context Window SupportWe implement and improve computation-communication overlapping, sequence parallelism, and communication compression to support up to 200K context length continue pretraining and finetuning. Our method to scale the context length to 200K is _solely_ based on engineering, that is to say, we do not modify the model architecture like sparse, local, or sliding window attention - the model remains using the full attention even the input is 200K. ## 5 Safety To enhance the model's trustworthiness and safety, we develop a full-stack Responsible AI Safety Engine (RAISE). RAISE ensures safe pretraining, alignment, and deployment. This section discusses our safety measures in the pretraining and alignment stages. Safety in PretrainingAligning with standard pretraining data safety practices [5; 58; 77], we build a set of filters based on heuristic rules, keyword matching, and learned classifiers to remove text containing personal identifiers and private data, and reduce sexual, violent, and extremist content. Safety in AlignmentInformed by existing research in [24; 35], we first build a comprehensive safety taxonomy. This taxonomy covers a broad spectrum of potential concerns, including environmental disharmony, superstitious, religious sensitivities, discriminatory practices, substance abuse, violent behavior, illegal activities, hate speech, ethical violations, privacy breaches, self-harm, sexually explicit content, mental health issues, and cybersecurity threats. We curated datasets reflecting these categories for a robust alignment, and mix them with our dialog SFT data. We also include a targeted set of prompts simulating attack scenarios in the alignment phase, which effectively improved the model's resilience against malicious use. ## 6 Evaluations Our evaluation demonstrates that the Yi model family achieves inspiring performance on a wide range of tasks and delivers close to GPT-3.5 user preference rate. We first report the base model performance on standard benchmarks, then we discuss the chat model performance and its user preference rate. ### Base Model Performance #### 6.1.1 Main Results Here we present the results for our base models and several other well-known base models across standard academic benchmarks. While benchmarking open-source models, we observed a disparity between the results generated by our pipeline and those reported in public sources. Upon conducting a more in-depth investigation of this difference, mostly because different models use different prompts, post-processing strategies, and sampling techniques. These differences may potentially induce significant variations in the outcomes. Our prompt and post-processing strategy remains consistent with the default settings of the original benchmarks[2; 4; 7; 8; 10; 11; 12; 27; 28; 42; 50; 61; 62; 63; 72; 74; 75; 89; 90]. We use greedy decoding without any post-processing for the generated content. For scores that were not reported publicly (or scores reported with different settings), we try to get results with our pipeline. For scores that can be found publicly, we directly report the existing numbers. We use the following benchmarks, largely following the practice of LLaMA 2 [77]: **Commonsense Reasoning:**: We included PIQA[4], SIQA[63], HellaSwag[89], WinoGrande [62], ARC[11], OpenBookQA(OBQA)[50], and CommonsenseQA(CSQA)[75] to assess common sense reasoning. CSQA was exclusively tested using a 7-shot setup, while all other tests were conducted with a 0-shot configuration. **Reading Comprehension:**: For reading comprehension, we report the 0-shot average on SQuAD[61], QuAC[8], and BoolQ[10]. **Math:**: We report the average of the GSM8K[12] (8 shot), and MATH[28] (4 shot) benchmarks with pass@1 accuracy without any specific prompting strategy (e.g. Chain-of-Thought prompting) and other ensemble technique (e.g., majority voting). **Code:**: We report the average pass@1 scores of our models on HumanEval[7] (Chen et al., 2021) and MBPP[2] (Austin et al., 2021). **Popular Aggregated Benchmark:**: We report the overall results for MMLU[27](5-shot), CMMLU[42] (5-shot), Gaokao-Bench[90] (5-shot), and BigBench[72] Hard (BBH[74]) (3-shot). By training on a significantly larger number of tokens (3.1T) compared to prior work (usually \(\leq 2\)T), we have observed a substantial performance gain across benchmarks, as shown in Table 2. However, it is important to note that there are still discernible disparities between our model and existing open-source and close-source models, particularly in tasks related to mathematics and coding. As performance in these domains can be significantly improved by continual pretraining and instruction fine-tuning, we have refrained from incorporating extensive mathematical and coding content in the pretraining corpus when making the initial design choices. We do plan to release models with enhanced math and coding capabilities in the future. #### 6.1.2 Discussions **Gain from Model Scale.**: We observe that Yi-34B has substantial performance improvement compared to Yi-6B, though they utilized the same pretrain corpora. Larger model size leads to higher performance gain on Code and Math benchmarks, referring to Tab. 3, compared to benchmarks focusing on Commonsense Reasoning, Reading Comprehension, or Knowledge. **Data Quality.**: Smaller models of higher quality pretrain data, like Yi-34B or Qwen-14B, usually demonstrate better performance than models of larger size but (presumably) lower quality data, such as Falcon-180B (though the focus of Falcon-180B might be more on the scaling side, which is definitely of important value on its own). **Gap between GPT-4 and Open-source LLMs.**: Based on Tab. 2, we note that open-source LLMs still lag behind the performance of GPT-4 and GPT-3.5 on various benchmarks. Yet representative bilingual LLMs, e.g. Qwen-14B and Yi-34B, can match or even surpass the performance of GPT-4 on Chinese knowledge related benchmarks, including C-Eval [31], CMMLU [42], and Gaokao [90]. However, there is still a huge gap between GPT-4 and open-source models on reasoning-related benchmarks like BBH [72], code (HumanEval), and math (MATH). \begin{table} \begin{tabular}{l c c c c c} \hline \hline **Model** & **Size** & ** GSM8k** & **MATH** & **Human-Eval pass@1** & **MBPP pass@1** \\ \hline **GPT-3.5** & - & 57.1 & 14.0 & 48.1 & 61.4 \\ **GPT-4** & - & **92.0** & **40.2** & **67.0** & **63.6** \\ \hline **Falcon** & 180B & 54.4 & - & 0.61 & 47.0 \\ \hline **Qwen** & 7B & 51.7 & 11.6 & 29.9 & 34.0 \\ & 14B & 61.3 & 24.8 & 32.3 & 48.9 \\ \hline **Baichuan 2** & 7B & 24.5 & 5.6 & 18.3 & 28.3 \\ & 13B & 22.1 & 10.1 & 20.7 & 26.1 \\ \hline **LLaMA 2** & 7B & 16.7 & 3.3 & 12.8 & 14.8 \\ & 34B & 42.2 & 6.2 & 22.6 & 33.0 \\ & 70B & 56.8 & 13.5 & 31.7 & 45.0 \\ \hline **Mistral** & 7B & 47.5 & 11.3 & 30.5 & 47.5 \\ \hline **InternLM** & 20B & 62.9 & 10.9 & 28.1 & 41.4 \\ \hline **Skywork** & 7B & 55.8 & 7.8 & 13.4 & 22.8 \\ \hline **Yi** & 6B & 32.5 & 4.6 & 15.9 & 26.3 \\ & 34B & 67.2 & 14.4 & 23.2 & 41.0 \\ \hline \hline \end{tabular} \end{table} Table 3: Comparison of models on GSM8k, MATH, Human-Eval, and MBPP. #### 6.1.3 In-Context Learning Study We further investigate the in-context learning capability, i.e., the capability of inferring the underlying function given the few-show input-output demonstrations. We consider the task of inferring the linear coefficient of a weighted sum. Specifically, define \(y=w_{1}x_{1}+w2x_{2}+...+w_{n}x_{n}\), our few-shot demonstration is \(x_{1},x_{2},...,x_{n},y\), and we ask the model to (implicitly) infer \(w_{1},w_{2},...,w_{n}\) by predicting the \(y\) given a new set of input \(x\). We use (a). the absolute difference between model prediction \(y\) and the ground truth \(y^{*}\), i.e., \(|y-y^{*}|\) as a continuous measure, and use (b). the exact match \(y==y^{*}\) as a discontinuous measure. We further note that most of the models perform reasonably well on addition and subtraction, so the ability to do arithmetic, as a confounding factor, can be ruled out. The results are shown in Figure 3. When setting the linear coefficients of be [1, -1], we see that Yi 34B and LLaMA-2 70B performs the best in-terms of answer exact match. If we increase the number of the linear coefficients to be [1, 1, 1, 1, 1], we observe the emergent behavior that only large models (LLaMA-2 70B and Mixtral) can achieve good scores on exact match, although the differences to target is more continuous. These observations give side evidence for Yi-34B's performance on in-context learning and indicates that further scaling may allow the model to infer more complicated functions by in-context learning. Figure 3: Evaluating language model’s in-context learning capability by inferring the linear coefficients of a weighted sum. Considering the discussions of whether emergent ability is an artifact of measurement [65], we use difference to the target (target number - model prediction) as a continuous measure, and exact match (target number == model prediction) as a discontinuous measure. A: when there is two linear coefficients, Yi-34B performs the best when measuring by the difference to the target number. B: increasing the number of linear coefficients to 5, only models that are large enough (LLaMA2 70B and Mixtral 8x7B) can achieve meaningful exact match, showing that in-context learning complex functions is an emergent ability. ### Chat Model Performance In this section, we report the automatic and human preference evaluation of the Chat Model. We use greedy decoding to generate responses. For the automatic evaluation benchmarks, we extract answers from the model's generated outputs and calculate accuracy. During the evaluation process, we observed that different prompts have varying influence on results. Therefore, for the same set of questions, we use identical prompts to evaluate all models, aiming to ensure as fair and unbiased results as possible. #### 6.2.1 Automatic Evaluations For automatic evaluation, we use the same benchmarks as is for the base model, detailed in Sec. 6.1.1. We use both zero-shot and few-shot methods but generally, zero-shot is more suitable for chat models. Our evaluation involves generating responses while following instructions explicitly or implicitly (such as the format in the few-shot examples). We then isolate relevant answers from the generated text. Unlike the base model, for the zero-shot evaluations on the GSM8K and BBH datasets, we employ the Chain-of-Thought (CoT) approach to guide the model in deliberation before reaching an answer. The results shown in Tab. 4 demonstrate the effectiveness of our chat models in understanding human instructions and generating appropriate instruction-following responses. We particularly highlight the 4-bit quantization results, as 4-bit quantization substantially reduces the memory requirement while the model performance nearly does not drop. This observation serve as the foundation of serving the model on consumer-grade devices. In line with Goodhart's principle, when a measurement metric becomes the target of our pursuit, it ceases to serve as a reliable standard of assessment. Consequently, the outcomes of our evaluations on benchmarks are exclusively employed for ensuring that our alignment training does not detrimentally impact the foundational knowledge and capabilities of the base model. We do not engage in targeted optimization of our chat model with the objective of enhancing benchmark performance. To further evaluate the generalizability of our model's capabilities, we conducted assessments of its mathematical computation proficiency by subjecting it to the 2023 Hungarian high school mathematics final exam questions, first proposed by the xAI Groke team then reproduced by Paster [55]. This evaluation was undertaken with the aim of determining whether our model exhibited signs of overfitting to training datasets that are mathematically oriented. The results in Fig. 4 show that Yi-34B-Chat performs inspiringly on both the GSM8K and the Hungarian mathematics exam. However, note that Yi-6B-Chat does not exhibit strong mathematical capabilities (on both GSM8K and the Hungarian mathematics exam). We speculate that smaller models may require more data to activate their corresponding abilities during the SFT stage. #### 6.2.2 Human Evaluations In this section we conducted an assessment of the model's conversational abilities, considering aspects to ensure its effectiveness and safety. We have compiled a collection of open-source evaluation datasets from the community, such as alpaca-eval[21], Belle-eval [88], and MT-bench[93]. Additionally, we have established our own helpful and harmless evaluation dataset by gathering and constructing data of varying difficulty levels, for the purpose of comprehensively assessing the conversational abilities of chat models. However, whether it is a public evaluation set or a self-built evaluation set, the evaluation results are strongly influenced by the assessment criteria and the design of the prompt. Our internal evaluation results may be unfair to other models, making it difficult to accurately represent the true capability level of our model. Therefore, here we only present external evaluation results to demonstrate the current conversational abilities of our chat model. We consider: (1). AlapcaEval1[44], which is designed to assess the English conversation capabilities of models by comparing the responses of a specified model to reference replies from Davinci003 [21] in order to calculate a win-rate; (2). LMSys2[93] Chatbot Arena, which showcases the responses of different models through a dialogue platform, then asks users to make selections based on their preferences, then computes the Elo score;(3). SuperClue3, on the other hand, is a leaderboard aimed at comprehensively evaluating the Chinese language capabilities of models. Footnote 3: [https://www.superclueai.com/](https://www.superclueai.com/) Tab. 5 presents the performance results of Yi-34B-Chat in the three third-party evaluations we consider, with the cutoff date for the results being December 21, 2023. The data demonstrates that, although there is still a gap compared to GPT-4, our model exhibits proficient bilingual (Chinese and English) dialogue capabilities and aligns well with user preferences. Additional comparative results of various models are accessible for review on the official website. \begin{table} \begin{tabular}{l c c c c c c} \hline \hline **Model** & **Size** & **MMLU** & **CMMLU** & **C-Eval(val)** & **TruthfulQA** & **BBH** & **GSM8K** \\ & \multicolumn{3}{c}{**0-shot / 5-shot 0-shot / 5-shot 0-shot / 5-shot**} & **0-shot & **0-shot / 3-shot** & **0-shot / 3-shot** & **0-shot / 4-shot** \\ \hline 1LaMA2-Chat & 13B & 50.9 / 47.3 & 27.5 / 35.1 & 27.9 / 35.9 & 36.8 & 32.9 / 58.2 & 36.9 / 2.7 \\ & 70B & 59.4 / 59.9 & 36.1 / 41.0 & 35.0 / 41.3 & 54.0 & 42.4 / 58.5 & 47.1 / 58.7 \\ \hline Baichuan2-Chat & 13B & 55.1 / 50.1 & 58.6 / 59.5 & 56.0 / 54.8 & 49.0 & 38.8 / 47.2 & 45.7 / 23.3 \\ \hline Qwen-Chat & 14B & 64.0 / 65.0 & 67.7 / 70.6 & 66.1 / 70.1 & 52.5 & 49.7 / 55.0 & 59.5 / 61.2 \\ \hline InterLM-Chat & 20B & 55.6 / 57.4 & 53.6 / 53.8 & 51.2 / 53.6 & 51.8 & 42.4 / 36.7 & 15.7 / 43.4 \\ \hline AquilaChat2 & 34B & 65.2 / 66.7 & 67.5 / 70.0 & **83.0 / 89.4** & **64.3** & 20.1 / 34.3 & 11.5 / 48.5 \\ \hline Yi-Chat & 6B & 58.2 / 61.0 & 69.4 / 74.7 & 68.8 / 74.2 & 50.6 & 39.7 / 47.2 & 38.4 / 44.9 \\ Yi-Chat-8bits(GPTQ) & 6B & 58.3 / 61.0 & 69.2 / 74.7 & 69.2 / 73.9 & 49.9 & 40.4 / 47.3 & 39.4 / 44.9 \\ Yi-Chat-4bits(AWQ) & 6B & 56.8 / 59.9 & 67.7 / 73.3 & 67.5 / 72.3 & 50.3 & 37.7 / 43.6 & 35.7 / 38.4 \\ \hline Yi-Chat & 34B & **67.6** / 73.5 & **79.1 / 81.3** & 77.0 / 78.5 & 62.4 & 51.4 / **71.7** & **71.7** / **76.0** \\ Yi-Chat-8bits(GPTQ) & 34B & 66.2 / **73.7** & 79.1 / 81.2 & 76.8 / 79.0 & 61.8 & **52.1** / 71.0 & 70.7 / 75.7 \\ Yi-Chat-4bits(AWQ) & 34B & 65.8 / 72.4 & 78.2 / 80.5 & 75.7 / 77.3 & 61.8 & 48.3 / 69.4 & 70.5 / 74.0 \\ \hline \hline \end{tabular} \end{table} Table 4: Overall performance on automatic benchmarks compared to open-source chat models. We highlight the 4-bit quantization results, as 4-bit quantization substantially reduces the memory requirement while the model performance nearly does not drop. This observation serve as the foundation of serving the model on consumer-grade devices, e.g., RTX4090. Figure 4: Yi’s result of Hungarian mathematics exam. We further demonstrate the data quality by comparing the speed of preference increase during data scaling. As is shown in Fig. 5, when compared with UltraChat [19] and its cleaned version UltraChat 200K, we see a clear tendency of performance improvements when scaling up the Yi data. ## 7 Capability Extension In this section, we discuss our post-training methods to extend the Yi base model to 200K long-context, equip it with visual understanding capability, and enhance the 6B model by depth upscaling. ### Long Context Modeling Our long-context solution consists of a continual pretraining and a finetuning phase, both are lightweight. We hold the basic hypothesis that the potential of utilizing information anywhere within the 200K input context is already exist in the base model (same as Fu et al. 22), the continue pretraining phase "unlocks" such capability, evidenced by a strong performance on Needle-in-a-Haystack test, then the finetuning phase further adapt the style of response to follow human instruction and preference. Continue PretrainingWe continue pretrain the full-attention model using sequence parallelism [43] and distributed attention. This is to say, we do not use any sparse or linear attention, but use a brute force implementation of the full attention. We continue pretrain the Yi 6B/ 34B base model on the data mixture of (1). original pretraining data, as is introduced in section 2; (2). length-upsampled long-context data, where the long documents are mostly from books; (3). multi-document question-answering synthetic data, where we construct QA pairs where the answer contains a recitation of the related paragraph before the answer. Our data approach mostly follows the data engineering practice in Fu et al. [22] and Yu et al. [87]. We continue pretrain the model on 5B tokens with 4M batch size, which translate to 100 optimization steps. Aligning with the concurrent work from Fu et al. [22], we observe that such light-weight continue pretraining is already able to enable a strong performance on Needle-in-a-Haystack test, as we will show in Figure 6. Supervised FinetuningWe mix our short-context SFT data with long-context document question-answering data. We use model-assisted automated methods (i.e., synthetic data) to construct document QA. Specifically, we randomly concatenate multiple documents into a sequence, sample one or more paragraphs from the long sequence, and ask a chat model to construct question and answer pairs Figure 5: SFT data scaling curve. Compared with UltraChat and its cleaned version UltraChat 200K, our SFT data demonstrates clear scaling advantages. We attribute its steep slope to the data quality. \begin{table} \begin{tabular}{l c c c c} \hline \hline **Model** & **Size** & **AlpacaEval** & **LMSys Chatbot Arena** & **SuperClue** \\ \hline GPT-4-Turbo & - & **97.7** & **1243** & **89.79** \\ GPT-3.5-Turbo & - & 89.37 & 1117 & 59.39 \\ LLaMA2-Chat & 70B & 92.66 & 1077 & - \\ Yi-Chat & 34B & 94.08 & 1110 & 71.87 \\ \hline \hline \end{tabular} \end{table} Table 5: Human evaluation comparison with other open-source chat models. [MISSING_PAGE_FAIL:15] Yi-VL models undergo a three-stage training process: **Stage 1:**: we train the parameters of the ViT and the projection module using an image resolution of \(224^{2}\). The training leverages a substantial dataset comprising \(100\) million image-text pairs from LAION-400M [66]. The primary objective is to enhance the ViT's knowledge acquisition within our specified architecture and to achieve better alignment between the ViT and the LLM. **Stage 2:**: we scale up the image resolution of ViT to \(448^{2}\), aiming to further boost the model's capability for discerning intricate visual details. The dataset used in this stage includes \(20\) million image-text pairs derived from LAION-400M. Additionally, we incorporate around \(4.8\) million image-text pairs from diverse sources, _e.g._, CLLaVA [45], LLaVAR [91], Flickr [85], VQAv2 [25], RefCOCO [37], Visual7w [95] and so on. **Stage 3:**: the parameters of the entire model are trained. The primary goal is to enhance the model's proficiency in multimodal chat interactions, thereby endowing it with the ability to seamlessly integrate and interpret visual and linguistic inputs. To this end, the training dataset encompasses a diverse range of sources, totalling approximately \(1\) million image-text pairs, including GQA [32], VizWiz VQA [26], TextCaps [71], OCR-VQA [51], Visual Genome [39], ShareGPT4V [6] and so on. To ensure data balancing, we impose a cap on the maximum data contribution from any single source, restricting it to no more than \(50,000\) pairs. In Stage 1 and 2, we set the global batch size, the learning rate, the gradient clip and the number of epoch to \(4096\), \(1\mathrm{e}{-4}\), \(0.5\) and \(1\), respectively. In Stage 3, these parameters are adjusted to \(256\), \(2\mathrm{e}{-5}\), \(1.0\) and \(2\). The training consumes \(128\) NVIDIA A100 GPUs. The total training time amounted to approximately \(3\) days for Yi-VL-6B and \(10\) days for Yi-VL-34B. Table 7 shows the MMMU test set leaderboard by Yi-VL's release. We note that this area is currently actively under research, aligning with the community's advances, we will continuously improve the update Yi-VL's performance. ### Depth Upscaling Recent studies on scaling laws [29; 30; 36] have underscored the predictable improvement in model performance with increases in computational budget, model size, and data size. Yet, identifying the most effective distribution of resources between model and data sizes upon expanding the computational budget remains a formidable challenge in the field of scaling laws. Additionally, research conducted by DeepSeek-AI et al. [17] has highlighted that the allocation of an increased computational budget towards model scaling should be proportional to the quality of the data available. In light of these insights, we propose a novel approach aimed at dynamically adjusting the resource allocation between data and model sizes through a series of staged training processes. This strategy iteratively fine-tunes the balance between data characteristics and model size according to scaling laws, enhancing both model training efficiency and performance. Figure 7: Architecture of Yi-VL models. Symbols are used to denote the training status of various modules at three training stages: a fire icon ( ) indicates the parameters of the module are trainable, while a snowflake icon () signifies that parameters are frozen. The image resolution used in ViT at each stage, either \(224^{2}\) or \(448^{2}\), is also marked. MethodFollowing the methodology outlined by Kim et al. [38], our goal is to upscale our Yi-6B base model, which has 32 layers, to a 9B model named the Yi-9B base model, featuring 48 layers, by duplicating the original 16 middle layers 12-28. Depth up-scaling involves expanding the base model's depth and subsequently continuing the pretraining phase for the enhanced model. Our investigations reveal that the decision on which layers to replicate could be informed by evaluating the cosine similarity scores between the inputs and outputs of each layer. Such an approach allows for targeted model scaling without necessitating additional pretraining, leading only to minimal performance impacts. This minimal impact on performance is attributed to the high cosine similarity, approaching one, between the inputs and outputs of the duplicated layers, as evidenced in Figure 8. This observation suggests that the replication of these layers does not significantly alter the output \begin{table} \begin{tabular}{l c c c c c c c} \hline \hline **Model** & **Overall** & **Art** & **Business** & **Science** & **Health** & **Society** & **Engineering** \\ \hline GPT-4V & 55.7 & 65.3 & 64.3 & 48.4 & 63.5 & 76.3 & 41.7 \\ Yi-VL-34B & 41.6 & 56.1 & 33.3 & 32.9 & 45.9 & 66.5 & 36.0 \\ Qwen-VL-PLUS & 40.8 & 59.9 & 34.5 & 32.8 & 43.7 & 65.5 & 32.9 \\ Marco-VL & 40.4 & 56.5 & 31.0 & 31.0 & 46.9 & 66.5 & 33.8 \\ Yi-VL-6B & 37.8 & 53.4 & 30.3 & 30.0 & 39.3 & 58.5 & 34.1 \\ InfMIM-Zephyr-7B & 35.5 & 50.0 & 29.6 & 28.2 & 37.5 & 54.6 & 31.1 \\ SVIT & 34.1 & 48.9 & 28.0 & 26.8 & 35.5 & 50.9 & 30.7 \\ Emu2-Chat & 34.1 & 50.6 & 27.7 & 28.0 & 32.4 & 50.3 & 31.3 \\ BLIP-2 FLAN-T5-XXL & 34.0 & 49.2 & 28.6 & 27.3 & 33.7 & 51.5 & 30.4 \\ InstructBLIP-T5-XXL & 33.8 & 48.5 & 30.6 & 27.6 & 33.6 & 49.8 & 29.4 \\ LLAVA-1.5-13B & 33.6 & 49.8 & 28.2 & 25.9 & 34.9 & 54.7 & 28.3 \\ Qwen-VL-7B-Chat & 32.9 & 47.7 & 29.8 & 25.6 & 33.6 & 45.3 & 30.2 \\ SPHINX* & 32.9 & 50.9 & 27.2 & 25.3 & 34.1 & 51.2 & 27.8 \\ mPLUG-OWL2 & 32.1 & 48.5 & 25.6 & 24.9 & 32.8 & 46.7 & 29.6 \\ BLIP-2 FLAN-T5-XL & 31.0 & 43.0 & 25.6 & 25.1 & 31.8 & 48.0 & 27.8 \\ InstructBLIP-T5-XL & 30.6 & 43.3 & 25.2 & 25.2 & 29.3 & 45.8 & 28.6 \\ CogVLM & 30.1 & 38.0 & 25.6 & 25.1 & 31.2 & 41.5 & 28.9 \\ \hline \hline \end{tabular} \end{table} Table 7: MMMU test set performance by the time of Yi-VL’s release. Figure 8: Input/output cosine similarity score of each token per layer for text ”Write a quiz about bits”. The cosine similarity scores of the 16 newly added layers(layers 28-44), as depicted in the lower figure, are observed to be nearly 1. logits produced by the original model. This method ensures the efficient scaling of the model by optimizing its architecture based on the internal processing dynamics of its layers. Continual TrainingThe dataset is composed of approximately 800 billion tokens across two stages, with around 70% having been recently collected and carefully selected. We have enhanced the code coverage in the final stage to improve code performance. To optimize the training process, we maintain a constant learning rate of 3e-5, and adopt a strategic approach to gradually increase the batch size from 4M tokens whenever the model's loss plateaued. This incremental adjustment of the batch size, alongside maintaining all other parameters in alignment with the established Yi-6B base model configuration, was instrumental in navigating the challenges of training at scale. The effectiveness of these strategies is demonstrated in Table 8, which details the Yi-9B base model's performance across a variety of benchmarks, including common sense, reasoning, knowledge, coding, and mathematics. It underscores the competitive advantages of Yi-9B base model in specific domains, illustrating the efficacy of our methodology in enhancing model performance by optimally adjusting the interplay between data characteristics and model size. ## 8 Final Discussions In this report, we discuss the full-stack development of the Yi language model family. Yi-34B achieves GPT-3.5 matching performance and is deployable (thank to the 4/8-bit quantization) on consumer-grade devices, making it an ideal model for local deployment. The key takeaways from the Yi pretraining procedure are about data quantity and quality: (1). training the model on a larger amount of data than the Chinchilla optimal delivers clear and consistent performance gain, which we highly recommend for all pretraining teams. Our model is trained on 3.1T tokens, yet we belive with larger amount of data, we can continue improve the model performance (i.e., the model have not saturated at 3.1T); (2). when it comes to the pretraining data quality, we believe the most critical two factors are the source of the data (e.g., whether the text is produced for professional usage or for casual social media posting) and the details of the data cleaning (e.g., the strength of filtering and deduplication). Since data cleaning is a very complicated pipeline and it is extremely difficult to conduct extensive grid-search styled optimizations, our current solution may still have room for improvements. The key takeaways from the Yi finetuning procedure is to heavily iterate on a small amount of data (\(\leq\) 10K), case by case, over multiple iterations, directly by the machine learning engineer, and improved from real user feedback. This approach clearly deviates from the instruction-scaling approach, initially introduced by the FLAN series [9] then followed by the UltraChat series [19]. As is demonstrated by our current results, the reasoning capability, which we view as the core capability for real-world deployment of language models, is strongly correlated with model scale when the amount of pretraining data is fixed. We believe that given our current results, continuing to scale up model parameters using thoroughly optimized data will lead to even stronger frontier models in our upcoming next versions. \begin{table} \begin{tabular}{l c c c c c c c c} \hline \hline **Model** & **Arc-C** & **HellaSwag** & **MMLU** & **Winogrande** & ** GSM8K** & **MATH** & **HumanEval** & **MBPP** \\ \hline **Vi-6B** & 50.3 & 74.4 & 63.2 & 71.3 & 32.5 & 4.6 & 15.9 & 26.3 \\ **Vi-9B Init** & 52.1 & 73.3 & 63.0 & 69.4 & 31.3 & 4.1 & 12.8 & 25.8 \\ **Vi-9B** & **55.6** & **76.4** & **68.4** & **73.0** & **52.3** & **15.9** & **39.0** & **54.4** \\ \hline \hline \end{tabular} \end{table} Table 8: Performance between Yi-6B and Yi-9B: Arc Challenge (25-shot), HellaSwag (10-shot) MMLU (5-shot), Winogrande (5-shot), GSM8K (5-shot), MATH (4-shot), HumanEval pass@1, MBPP pass@1(3-shot). Yi-9B Init is just depthwise upscaling from Yi-6B by duplicating layers 12-28 without further training. Author List and Contributions Our team members contribute to the development of Yi from the following perspectives: * Frontier Research * Machine Learning Infrastructure * Pretraining * Finetuning and AI Alignment * Multimodal * Safety and Responsible AI * Deployment We list our team members in alphabetical order. All authors contributed equally to this work. * Alex Young * Bei Chen * Chao Li * Chengen Huang * Ge Zhang * Guanwei Zhang * Heng Li * Jiangcheng Zhu * Jianqun Chen * Jing Chang * Kaidong Yu * Peng Liu * Qiang Liu * Shawn Yue * Senbin Yang * Multimodal * Safety and Responsible AI * Deployment * Shiming Yang * Tao Yu * Wen Xie * Wenhao Huang * Xiaohui Hu * Xiaoyi Ren * Xinyao Niu * Pengcheng Nie * Yuchi Xu * Yudong Liu * Yue Wang * Yuxuan Cai * Zhenyu Gu * Zhiyuan Liu * Zonghong Dai ## References * [1] Joshua Ainslie, James Lee-Thorp, Michiel de Jong, Yury Zemlyanskiy, Federico Lebron, and Sumit Sanghai. GQA: Training Generalized Multi-Query Transformer Models from Multi-Head Checkpoints. _arXiv preprint arXiv:2305.13245_, 2023. * [2] Jacob Austin, Augustus Odena, Maxwell Nye, Maarten Bosma, Henryk Michalewski, David Dohan, Ellen Jiang, Carrie Cai, Michael Terry, Quoc Le, et al. Program Synthesis With ILarge Language Models. _arXiv preprint arXiv:2108.07732_, 2021. * [3] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, Binyuan Hui, Luo Ji, Mei Li, Junyang Lin, Runji Lin, Dayiheng Liu, Gao Liu, Chengqiang Lu, Keming Lu, Jianxin Ma, Rui Men, Xingzhang Ren, Xuancheng Ren, Chuanqi Tan, Sinan Tan, Jianhong Tu, Peng Wang, Shijie Wang, Wei Wang, Shengguang Wu, Benfeng Xu, Jin Xu, An Yang, Hao Yang, Jian Yang, Shusheng Yang, Yang Yao, Bowen Yu, Hongyi Yuan, Zheng Yuan, Jianwei Zhang, Xingxuan Zhang, Yichang Zhang, Zhenru Zhang, Chang Zhou, Jingren Zhou, Xiaohuan Zhou, and Tianhang Zhu. Qwen Technical Report. 09 2023. URL [https://arxiv.org/pdf/2309.16609.pdf](https://arxiv.org/pdf/2309.16609.pdf). * [4] Yonatan Bisk, Rowan Zellers, Ronan Le Bras, Jianfeng Gao, and Yejin Choi. PIQA: Reasoning about Physical Commonsense in Natural Language. _ArXiv_, abs/1911.11641, 2019. URL [https://api.semanticscholar.org/CorpusID:208290939](https://api.semanticscholar.org/CorpusID:208290939). * [5] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. _Advances in neural information processing systems_, 33:1877-1901, 2020. * [6] Lin Chen, Jisong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v: Improving large multi-modal models with better captions. _arXiv preprint arXiv:2311.12793_, 2023. * [7] Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harrison Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, Alex Ray, Raul Puri, Gretchen Krueger, Michael Petrov, Heidy Khaaf, Girish Sastry, Pamela Mishkin, Brooke Chan, Scott Gray, Nick Ryder, Mikhail Pavlov, Alethea Power, Lukasz Kaiser, Mohammad Bavarian, Clemens Winter, Philippe Tillet, Felipe Petroski Such, Dave Cummings, Matthias Plappert, Fotios Chantzis, Elizabeth Barnes, Ariel Herbert-Voss, William Hebgen Guss, Alex Nichol, Alex Paino, Nikolas Tezak, Jie Tang, Igor Babuschkin, Suchir Balaji, Shantanu Jain, William Saunders, Christopher Hesse, Andrew N. Carr, Jan Leike, Joshua Achiam, Vedant Misra, Evan Morikawa, Alec Radford, Matthew Knight, Miles Brundage, Mira Murati, Katie Mayer, Peter Welinder, Bob McGrew, Dario Amodei, Sam McCandlish, Ilya Sutskever, and Wojciech Zaremba. Evaluating Large Language Models Trained on Code. _CoRR_, abs/2107.03374, 2021. URL [https://arxiv.org/abs/2107.03374](https://arxiv.org/abs/2107.03374). * [8] Eunsol Choi, He He, Mohit Iyyer, Mark Yatskar, Wen tau Yih, Yejin Choi, Percy Liang, and Luke Zettlemoyer. QuAC : Question Answering in Context, 2018. * [9] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instruction-finetuned language models. _arXiv preprint arXiv:2210.11416_, 2022. * [10] Christopher Clark, Kenton Lee, Ming-Wei Chang, Tom Kwiatkowski, Michael Collins, and Kristina Toutanova. BoolQ: Exploring the Surprising Difficulty of Natural Yes/No Questions, 2019. * [11] Peter Clark, Isaac Cowhey, Oren Etzioni, Tushar Khot, Ashish Sabharwal, Carissa Schoenick, and Oyvind Tafjord. Think you have Solved Question Answering? Try ARC, the AI2 Reasoning Challenge, 2018. * [12] Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, et al. Training Verifiers to Solve Math Word Problems. _arXiv preprint arXiv:2110.14168_, 2021. * [13] Together Computer. Redpajama: an open dataset for training large language models, 2023. URL [https://github.com/togethercomputer/RedPajama-Data](https://github.com/togethercomputer/RedPajama-Data). * [14] Tri Dao. FlashAttention-2: Faster attention with better parallelism and work partitioning. 2023. * [15] Tri Dao, Daniel Y. Fu, Stefano Ermon, Atri Rudra, and Christopher Re. FlashAttention: Fast and memory-efficient exact attention with IO-awareness. In _Advances in Neural Information Processing Systems_, 2022. * [16] Michiel de Jong, Yury Zemlyanskiy, Joshua Ainslie, Nicholas FitzGerald, Sumit Sanghai, Fei Sha, and William Cohen. FiDO: Fusion-in-Decoder Optimized for Stronger Performance and Faster Inference. _arXiv preprint arXiv:2212.08153_, 2022. * [17] DeepSeek-AI, :, Xiao Bi, Deli Chen, Guanting Chen, Shanhuang Chen, Damai Dai, Chengqi Deng, Honghui Ding, Kai Dong, Qiushi Du, Zhe Fu, Huazuo Gao, Kaige Gao, Wenjun Gao, Ruiqi Ge, Kang Guan, Daya Guo, Jianzhong Guo, Guangbo Hao, Zhewen Hao, Ying He, Wenjie Hu, Panpan Huang, Erhang Li, Guowei Li, Jiashi Li, Yao Li, Y. K. Li, Wenfeng Liang, Fangyun Lin, A. X. Liu, Bo Liu, Wen Liu, Xiaodong Liu, Xin Liu, Yiyuan Liu, Haoyu Lu, Shanghao Lu, Fuli Luo, Shirong Ma, Xiaotao Nie, Tian Pei, Yishi Piao, Junjie Qiu, Hui Qu, Tongzheng Ren, Zehui Ren, Chong Ruan, Zhangli Sha, Zhihong Shao, Junxiao Song, Xuecheng Su, Jingxiang Sun, Yaofeng Sun, Minghui Tang, Bingxuan Wang, Peiyi Wang, Shiyu Wang, Yaohui Wang, Yongji Wang, Tong Wu, Y. Wu, Xin Xie, Zhenda Xie, Ziwei Xie, Yiliang Xiong, Hanwei Xu, R. X. Xu, Yanhong Xu, Dejian Yang, Yuxiang You, Shuiping Yu, Xingkai Yu, B. Zhang, Haowei Zhang, Lecong Zhang, Liyue Zhang, Mingchuan Zhang, Minghua Zhang, Wentao Zhang, Yichao Zhang, Chenggang Zhao, Yao Zhao, Shangyan Zhou, Shunfeng Zhou, Qihao Zhu, and Yuheng Zou. Deepseek llm: Scaling open-source language models with longtermism. 2024. * [18] Tim Dettmers, Mike Lewis, Younes Belkada, and Luke Zettlemoyer. Llm: int8 (): 8-bit matrix multiplication for transformers at scale. _arXiv preprint arXiv:2208.07339_, 2022. * [19] Ning Ding, Yulin Chen, Bokai Xu, Yujia Qin, Zhi Zheng, Shengding Hu, Zhiyuan Liu, Maosong Sun, and Bowen Zhou. Enhancing chat language models by scaling high-quality instructional conversations. _arXiv preprint arXiv:2305.14233_, 2023. * [20] Guanting Dong, Hongyi Yuan, Keming Lu, Chengpeng Li, Mingfeng Xue, Dayiheng Liu, Wei Wang, Zheng Yuan, Chang Zhou, and Jingren Zhou. How abilities in large language models are affected by supervised fine-tuning data composition, 2023. * [21] Yann Dubois, Xuechen Li, Rohan Taori, Tianyi Zhang, Ishaan Gulrajani, Jimmy Ba, Carlos Guestrin, Percy Liang, and Tatsunori B. Hashimoto. Alpacafarm: A simulation framework for methods that learn from human feedback, 2023. * [22] Yao Fu, Rameswar Panda, Xinyao Niu, Xiang Yue, Hannaneh Hajishirzi, Yoon Kim, and Hao Peng. Data engineering for scaling language models to 128k context. _arXiv preprint arXiv:2402.10171_, 2024. * [23] Gemini Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. Gemini: A family of highly capable multimodal models. _arXiv preprint arXiv:2312.11805_, 2023. * [24] Amelia Glaese, Nat McAleese, Maja Trebacz, John Aslanides, Vlad Firoiu, Timo Ewalds, Maribeth Rauh, Laura Weidinger, Martin Chadwick, Phoebe Thacker, et al. Improving alignment of dialogue agents via targeted human judgements. _arXiv preprint arXiv:2209.14375_, 2022. * [25] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In _Proceedings of the IEEE conference on computer vision and pattern recognition_, pages 6904-6913, 2017. * [26] Danna Gurari, Qing Li, Abigale J Stangl, Anhong Guo, Chi Lin, Kristen Grauman, Jiebo Luo, and Jeffrey P Bigham. Vizwiz grand challenge: Answering visual questions from blind people. In _Proceedings of the IEEE conference on computer vision and pattern recognition_, pages 3608-3617, 2018. * [27] Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. Measuring Massive Multitask Language Understanding. _CoRR_, abs/2009.03300, 2020. URL [https://arxiv.org/abs/2009.03300](https://arxiv.org/abs/2009.03300). * [28] Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. Measuring Mathematical Problem Solving With the MATH Dataset. _arXiv preprint arXiv:2103.03874_, 2021. * [29] Tom Henighan, Jared Kaplan, Mor Katz, Mark Chen, Christopher Hesse, Jacob Jackson, Heewowo Jun, Tom B. Brown, Prafulla Dhariwal, Scott Gray, Chris Hallacy, Benjamin Mann, Alec Radford, Aditya Ramesh, Nick Ryder, Daniel M. Ziegler, John Schulman, Dario Amodei, and Sam McCandlish. Scaling laws for autoregressive generative modeling. 2020. * [30] Jordan Hoffmann, Sebastian Borgeaud, Arthur Mensch, Elena Buchatskaya, Trevor Cai, Eliza Rutherford, Diego de Las Casas, Lisa Anne Hendricks, Johannes Welbl, Aidan Clark, et al. Training compute-optimal large language models. _arXiv preprint arXiv:2203.15556_, 2022. * [31] Yuzhen Huang, Yuzhuo Bai, Zhihao Zhu, Junlei Zhang, Jinghan Zhang, Tangjun Su, Junteng Liu, Chuancheng Lv, Yikai Zhang, Jiayi Lei, et al. C-eval: A multi-level multi-discipline chinese evaluation suite for foundation models. _arXiv preprint arXiv:2305.08322_, 2023. * [32] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In _Proceedings of the IEEE/CVF conference on computer vision and pattern recognition_, pages 6700-6709, 2019. * [33] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, July 2021. URL [https://doi.org/10.5281/zenodo.5143773](https://doi.org/10.5281/zenodo.5143773). * [34] Neel Jain, Ping-yeh Chiang, Yuxin Wen, John Kirchenbauer, Hong-Min Chu, Gowthami Sompalli, Brian R Bartoldson, Bhavya Kailkhura, Avi Schwarzschild, Aniruddha Saha, et al. Neptune: Noisy embeddings improve instruction finetuning. _arXiv preprint arXiv:2310.05914_, 2023. * [35] Jiaming Ji, Mickel Liu, Juntao Dai, Xuehai Pan, Chi Zhang, Ce Bian, Chi Zhang, Ruiyang Sun, Yizhou Wang, and Yaodong Yang. Beavertails: Towards improved safety alignment of llvm via a human-preference dataset, 2023. * [36] Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B. Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models. 2020. * [37] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In _Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP)_, pages 787-798, 2014. * [38] Dahyun Kim, Chanjun Park, Sanghoon Kim, Wonsung Lee, Wonho Song, Yunsu Kim, Hyeonwoo Kim, Yungi Kim, Hyeonju Lee, Jihoo Kim, Changbae Ahn, Seonghoon Yang, Sukyung Lee, Hyunbyung Park, Gyoungjin Gim, Mikyoung Cha, Hwalsuk Lee, and Sunghun Kim. Solar 10.7b: Scaling large language models with simple yet effective depth up-scaling. 2023. * [39] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. _International journal of computer vision_, 123:32-73, 2017. * [40] Taku Kudo and John Richardson. SentencePiece: A Simple and Language Independent Subword Tokenizer and Detokenizer for Neural Text Processing. _arXiv preprint arXiv:1808.06226_, 2018. * [41] Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Zhang, and Ion Stoica. Efficient Memory Management for Large Language Model Serving with PagedAttention. _arXiv preprint arXiv:2309.06180_, 2023. * [42] Haonan Li, Yixuan Zhang, Fajri Koto, Yifei Yang, Hai Zhao, Yeyun Gong, Nan Duan, and Timothy Baldwin. CMMLU: Measuring Massive Multitask Language Understanding in Chinese. _arXiv preprint arXiv:2306.09212_, 2023. * [43] Shenggui Li, Fuzhao Xue, Chaitanya Baranwal, Yongbin Li, and Yang You. Sequence parallelism: Long sequence training from system perspective. _arXiv preprint arXiv:2105.13120_, 2021. * [44] Xuechen Li, Tianyi Zhang, Yann Dubois, Rohan Taori, Ishaan Gulrajani, Carlos Guestrin, Percy Liang, and Tatsunori B. Hashimoto. Alpaceval: An automatic evaluator of instruction-following models. [https://github.com/tatsu-lab/alpaca_eval](https://github.com/tatsu-lab/alpaca_eval), 2023. * [45] LinkSoul-AI. Chinese llava. [https://github.com/LinkSoul-AI/Chinese-LLAVA](https://github.com/LinkSoul-AI/Chinese-LLAVA), 2023. * [46] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. _arXiv preprint arXiv:2310.03744_, 2023. * [47] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. _arXiv preprint arXiv:2304.08485_, 2023. * [48] Wei Liu, Weihao Zeng, Keqing He, Yong Jiang, and Junxian He. What makes good data for alignment? a comprehensive study of automatic data selection in instruction tuning. _arXiv preprint arXiv:2312.15685_, 2023. * [49] Keming Lu, Hongyi Yuan, Zheng Yuan, Runji Lin, Junyang Lin, Chuanqi Tan, Chang Zhou, and Jingren Zhou. #instag: Instruction tagging for analyzing supervised fine-tuning of large language models, 2023. * [50] Todor Mihaylov, Peter Clark, Tushar Khot, and Ashish Sabharwal. Can a Suit of Armor Conduct Electricity? A New Dataset for Open Book Question Answering, 2018. * [51] Anand Mishra, Shashank Shekhar, Ajeet Kumar Singh, and Anirban Chakraborty. Ocr-vqa: Visual question answering by reading text in images. In _2019 international conference on document analysis and recognition (ICDAR)_, pages 947-952. IEEE, 2019. * [52] Thuat Nguyen, Chien Van Nguyen, Viet Dac Lai, Hieu Man, Nghia Trung Ngo, Franck Dernoncourt, Ryan A Rossi, and Thien Huu Nguyen. CulturaX: A Cleaned, Enormous, and Multilingual Dataset for Large Language Models in 167 Languages. _arXiv preprint arXiv:2309.09400_, 2023. * [53] OpenAI. ChatML, 2022. URL [https://github.com/openai/openai-python/blob/e389823ba013a24b4c32ce38fa0bd87e6bccae94/chatml.md](https://github.com/openai/openai-python/blob/e389823ba013a24b4c32ce38fa0bd87e6bccae94/chatml.md). * [54] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training Language Models to Follow Instructions with Human Feedback. _Advances in Neural Information Processing Systems_, 35:27730-27744, 2022. * [55] Keiran Paster. Testing language models on a held-out high school national finals exam. [https://huggingface.co/datasets/keirp/hungarian_national_hs_finals_exam](https://huggingface.co/datasets/keirp/hungarian_national_hs_finals_exam), 2023. * [56] Guilherme Penedo, Quentin Malartic, Daniel Hesslow, Ruxandra Cojocaru, Alessandro Cappelli, Hamza Alobeidli, Baptiste Pannier, Ebtesam Almazrouei, and Julien Launay. The RefinedWeb Dataset for Falcon LLM: Outperforming Curated Corpora with Web Data, and Web Data Only, 2023. * [57] Reiner Pope, Sholto Douglas, Aakanksha Chowdhery, Jacob Devlin, James Bradbury, Jonathan Heek, Kefan Xiao, Shivani Agrawal, and Jeff Dean. Efficiently Scaling Transformer Inference. _Proceedings of Machine Learning and Systems_, 5, 2023. * [58] Jack W Rae, Sebastian Borgeaud, Trevor Cai, Katie Millican, Jordan Hoffmann, Francis Song, John Aslanides, Sarah Henderson, Roman Ring, Susannah Young, et al. Scaling Language Models: Methods, Analysis & Insights from Training Gopher. _arXiv preprint arXiv:2112.11446_, 2021. * [59] Rafael Rafailov, Archit Sharma, Eric Mitchell, Stefano Ermon, Christopher D Manning, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. _arXiv preprint arXiv:2305.18290_, 2023. * [60] Samyam Rajbhandari, Jeff Rasley, Olatunji Ruwase, and Yuxiong He. ZeRO: Memory Optimizations Toward Training Trillion Parameter Models. In _SC20: International Conference for High Performance Computing, Networking, Storage and Analysis_, pages 1-16. IEEE, 2020. * [61] Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. SQuAD: 100,000+ Questions for Machine Comprehension of Text, 2016. * [62] Keisuke Sakaguchi, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. WinoGrande: An Adversarial Winograd Schema Challenge at Scale, 2019. * [63] Maarten Sap, Hannah Rashkin, Derek Chen, Ronan LeBras, and Yejin Choi. SocialIQA: Commonsense Reasoning about Social Interactions, 2019. * [64] Nikhil Sardana and Jonathan Frankle. Beyond chinchilla-optimal: Accounting for inference in language model scaling laws. _arXiv preprint arXiv:2401.00448_, 2023. * [65] Rylan Schaeffer, Brando Miranda, and Sanmi Koyejo. Are emergent abilities of large language models a mirage? _Advances in Neural Information Processing Systems_, 36, 2024. * [66] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. _arXiv preprint arXiv:2111.02114_, 2021. * [67] Noam Shazeer. Fast Transformer Decoding: One Write-Head is All You Need. _arXiv preprint arXiv:1911.02150_, 2019. * [68] Noam Shazeer. GLU Variants Improve Transformer. _arXiv preprint arXiv:2002.05202_, 2020. * [69] Yusuxke Shibata, Takuya Kida, Shuichi Fukamachi, Masayuki Takeda, Ayumi Shinohara, Takeshi Shinohara, and Setsuo Arikawa. Byte Pair Encoding: A Text Compression Scheme That Accelerates Pattern Matching. Technical report, Technical Report DOI-TR-161, Department of Informatics, Kyushu University, 1999. * [70] Mohammad Shoeybi, Mostofa Patwary, Raul Puri, Patrick LeGresley, Jared Casper, and Bryan Catanzaro. Megatron-LM: Training Multi-Billion Parameter Language Models Using Model Parallelism. _arXiv preprint arXiv:1909.08053_, 2019. * [71] Oleksii Sidorov, Ronghang Hu, Marcus Rohrbach, and Amanpreet Singh. Textcaps: a dataset for image captioning with reading comprehension. In _Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16_, pages 742-758. Springer, 2020. * [72] Aarohi Srivastava, Abhinav Rastogi, Abhishek Rao, Abu Awal Md Shoeb, Abubakar Abid, Adam Fisch, Adam R. Brown, Adam Santoro, Aditya Gupta, Adria Garriga-Alonso, Agnieszka Kluska, Aitor Lewkowycz, Akshat Agarwal, Alethea Power, Alex Ray, Alex Warstadt, Alexander W. Kocurek, Ali Safaya, Ali Tazarv, Alice Xiang, Alicia Parrish, Allen Nie, Aman Hussain, Amanda Askell, Amanda Dsouza, Ambrose Slone, Ameet Rahane, Anantharaman S. Iyer, Anders Andreassen, Andrea Madotto, Andrea Santilli, Andreas Stuhlmuller, Andrew Dai, Andrew La, Andrew Lampinen, Andy Zou, Angela Jiang, Angelica Chen, Anh Vuong, Animesh Gupta, Anna Gottardi, Antonio Norelli, Anu Venkatesh, Arash Gholamidavoodi, Arfa Tabassum, Arul Menezes, Arun Kirubarajan, Asher Mullokandov, Ashish Sabharwal, Austin Herrick, Avia Erfat, Aykat Erdem, Ayla Karakas, B. Ryan Roberts, Bao Sheng Loe, Barret Zoph, Bartomiej Bojanowski, Bathan Ozyurt, Behnam Hedayatnia, Behnam Neyshabur, Benjamin Inden, Benno Stein, Berk Ekmekci, Bill Yuchen Lin, Blake Howald, Bryan Orinion, Cameron Diao, Cameron Dour, Catherine Stinson, Cedrick Argueta, Cesar Ferri Ramirez, Chandan Singh, Charles Rathkopf, Chenlin Meng, Chitta Baral, Chiyu Wu, Chris Callison-Burch, Chris Waites, Christian Voigt, Christopher D. Manning, Christopher Potts, Cindy Ramirez, Clara E. Rivera, Clemencia Siro, Colin Raffel, Courtney Ashcraft, Cristina Garbacea, Dami Sileo, Dan Garrette, Dan Hendrycks, Dan Kilman, Dan Roth, Daniel Freeman, Daniel Khashabi,Daniel Levy, Daniel Mosegui Gonzalez, Danielle Perszyk, Danny Hernandez, Danqi Chen, and Daphne Ippolito et al. (351 additional authors not shown). Beyond the Imitation Game: Quantifying and Extrapolating the Capabilities of Language Models, 2023. * Su et al. (2021) Jianlin Su, Yu Lu, Shengfeng Pan, Ahmed Murtadha, Bo Wen, and Yunfeng Liu. Roformer: Enhanced Transformer with Rotary Position Embedding. _arXiv preprint arXiv:2104.09864_, 2021. * Suzgun et al. (2022) Mirac Suzgun, Nathan Scales, Nathanael Scharli, Sebastian Gehrmann, Yi Tay, Hyung Won Chung, Aakanksha Chowdhery, Quoc V Le, Ed H Chi, Denny Zhou, et al. Challenging BigBench Tasks and Whether Chain-of-Thought can Solve Them. _arXiv preprint arXiv:2210.09261_, 2022. * Talmor et al. (2019) Alon Talmor, Jonathan Herzig, Nicholas Lourie, and Jonathan Berant. CommonsenseQA: A Question Answering Challenge Targeting Commonsense Knowledge, 2019. * Touvron et al. (2023) Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothe Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. _arXiv preprint arXiv:2302.13971_, 2023. * Touvron et al. (2021) Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. Llama 2: Open Foundation and Fine-Tuned Chat Models, 2023. * Vaswani et al. (2017) Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention Is All You Need. _Advances in Neural Information Processing Systems_, 06 2017. URL [https://arxiv.org/pdf/1706.03762.pdf](https://arxiv.org/pdf/1706.03762.pdf). * Wenzek et al. (2019) Guillaume Wenzek, Marie-Anne Lachaux, Alexis Conneau, Vishrav Chaudhary, Francisco Guzman, Armand Joulin, and Edouard Grave. CCNet: Extracting High Quality Monolingual Datasets from Web Crawl Data. _arXiv preprint arXiv:1911.00359_, 11 2019. URL [https://arxiv.org/pdf/1911.00359.pdf](https://arxiv.org/pdf/1911.00359.pdf). * Wenzek et al. (2019) Guillaume Wenzek, Marie-Anne Lachaux, Alexis Conneau, Vishrav Chaudhary, Francisco Guzman, Armand Joulin, and Edouard Grave. CCNet: Extracting High Quality Monolingual Datasets from Web Crawl Data. _arXiv preprint arXiv:1911.00359_, 2019. * Wu et al. (2023) Xiaoxia Wu, Cheng Li, Reza Yazdani Aminabadi, Zhewei Yao, and Yuxiong He. Understanding int4 quantization for transformer models: Latency speedup, composability, and failure cases. _arXiv preprint arXiv:2301.12017_, 2023. * Xiong et al. (2023) Wenhan Xiong, Jingyu Liu, Igor Molybog, Hejia Zhang, Prajwal Bhargava, Rui Hou, Louis Martin, Rashi Rungta, Karthik Abinav Sankararaman, Barlas Oguz, et al. Effective long-context scaling of foundation models. _arXiv preprint arXiv:2309.16039_, 2023. * Xu et al. (2023) Can Xu, Qingfeng Sun, Kai Zheng, Xiubo Geng, Pu Zhao, Jiazhan Feng, Chongyang Tao, and Daxin Jiang. Wizardlm: Empowering large language models to follow complex instructions. _arXiv preprint arXiv:2304.12244_, 2023. * Yang et al. (2021) Aiyuan Yang, Bin Xiao, Bingning Wang, Borong Zhang, Ce Bian, Chao Yin, Chenxu Lv, Da Pan, Dian Wang, Dong Yan, Fan Yang, Fei Deng, Feng Wang, Feng Liu, Guangwei Ai, Guosheng Dong, Haizhou Zhao, Hang Xu, Haoze Sun, Hongda Zhang, Hui Liu, Jiaming Ji, Jian Xie, JunTao Dai, Kun Fang, Lei Su, Liang Song, Lifeng Liu, Liyun Ru, Luyao Ma, MangWang, Mickel Liu, MingAn Lin, Nuolan Nie, Peidong Guo, Ruiyang Sun, Tao Zhang, Tianpeng Li, Tianyu Li, Wei Cheng, Weipeng Chen, Xiangrong Zeng, Xiaochuan Wang, Xiaoxi Chen, Xin Men, Xin Yu, Xuehai Pan, Yanjun Shen, Yiding Wang, Yiyu Li, Youxin Jiang, Yuchen Gao, Yupeng Zhang, Zenan Zhou, and Zhiying Wu. Baichuan 2: Open Large-scale Language Models. 09 2023. URL [https://arxiv.org/pdf/2309.10305.pdf](https://arxiv.org/pdf/2309.10305.pdf). * Young et al. [2014] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. _Transactions of the Association for Computational Linguistics_, 2:67-78, 2014. * Yu et al. [2022] Gyeong-In Yu, Joo Seong Jeong, Geon-Woo Kim, Soojeong Kim, and Byung-Gon Chun. Orca: A Distributed Serving System for Transformer-Based Generative Models. In _16th USENIX Symposium on Operating Systems Design and Implementation (OSDI 22)_, pages 521-538, 2022. * Yu et al. [2023] Yijiong Yu, Zhe Zhou, Zhixiao Qi, and Yongfeng Huang. Paraphrasing the original text makes high accuracy long-context qa. _arXiv preprint arXiv:2312.11193_, 2023. * Yunjie et al. [2023] Ji Yunjie, Deng Yong, Gong Yan, Peng Yiping, Niu Qiang, Zhang Lei, Ma Baochang, and Li Xiangang. Exploring the impact of instruction data scaling on large language models: An empirical study on real-world use cases. _arXiv preprint arXiv:2303.14742_, 2023. * Zellers et al. [2019] Rowan Zellers, Ari Holtzman, Yonatan Bisk, Ali Farhadi, and Yejin Choi. HellaSwag: Can a Machine Really Finish Your Sentence?, 2019. * Zhang et al. [2023] Xiaotian Zhang, Chunyang Li, Yi Zong, Zhengyu Ying, Liang He, and Xipeng Qiu. Evaluating the Performance of Large Language Models on GAOKAO Benchmark. _arXiv preprint arXiv:2305.12474_, 2023. * Zhang et al. [2023] Yanzhe Zhang, Ruiyi Zhang, Jiuxiang Gu, Yufan Zhou, Nedim Lipka, Diyi Yang, and Tong Sun. Llavar: Enhanced visual instruction tuning for text-rich image understanding. _arXiv preprint arXiv:2306.17107_, 2023. * Zheng et al. [2023] Huaixiu Steven Zheng, Swaroop Mishra, Xinyun Chen, Heng-Tze Cheng, Ed H. Chi, Quoc V Le, and Denny Zhou. Take a step back: Evoking reasoning via abstraction in large language models, 2023. * Zheng et al. [2023] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric. P Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. Judging llm-as-a-judge with mt-bench and chatbot arena, 2023. * Zhou et al. [2023] Chunting Zhou, Pengfei Liu, Puxin Xu, Srini Iyer, Jiao Sun, Yuning Mao, Xuezhe Ma, Avia Efrat, Ping Yu, Lili Yu, Susan Zhang, Gargi Ghosh, Mike Lewis, Luke Zettlemoyer, and Omer Levy. Lima: Less is more for alignment, 2023. * Zhu et al. [2016] Yuke Zhu, Oliver Groth, Michael Bernstein, and Li Fei-Fei. Visual7w: Grounded question answering in images. In _Proceedings of the IEEE conference on computer vision and pattern recognition_, pages 4995-5004, 2016.

# _WizardCoder_: Empowering Code Large Language Models with Evol-Instruct Ziyang Luo\({}^{2}\)1 Can Xu\({}^{1}\)1 Pu Zhao\({}^{1}\) Qingfeng Sun\({}^{1}\) Xiubo Geng\({}^{1}\) **Wenxiang Hu\({}^{1}\) Chongyang Tao\({}^{1}\) Jing Ma\({}^{2}\) Qingwei Lin\({}^{1}\) Daxin Jiang\({}^{1}\)2 \({}^{1}\)Microsoft** \({}^{2}\)Hong Kong Baptist University {caxu,puzhao,qins,xigeng,wenxh,chongyang.tao,qlin,djiang}@microsoft.com {cszyluo, majing}@comp.hkbu.edu.hk Equal contribution. Work done during the internship at Microsoft.Corresponding author. Footnote 1: Corresponding author. ###### Abstract Code Large Language Models (Code LLMs), such as StarCoder, have demonstrated exceptional performance in code-related tasks. However, most existing models are solely pre-trained on extensive raw code data without instruction fine-tuning. In this paper, we introduce _WizardCoder_, which empowers Code LLMs with complex instruction fine-tuning, by adapting the _Evol-Instruct_ method to the domain of code. Through comprehensive experiments on four prominent code generation benchmarks, namely HumanEval, HumanEval+, MBPP, and DS-1000, we unveil the exceptional capabilities of our model. It surpasses all other open-source Code LLMs by a substantial margin. Moreover, our model even outperforms the largest closed LLMs, Anthropic's Claude and Google's Bard, on HumanEval and HumanEval+. Our code, model weights, and data are public at [https://github.com/nlpxucan/WizardLM](https://github.com/nlpxucan/WizardLM). ## 1 Introduction Recently, Large Language Models (LLMs) [1; 2; 3; 4; 5; 6; 7; 8; 9] have garnered significant attention and demonstrated impressive success. Notably, OpenAI's ChatGPT stands out as a prominent example. Leveraging extensive pre-training on vast amounts of internet data and further fine-tuning with detailed instruction data [10], these models have achieved state-of-the-art (SOTA) zero-shot performance across diverse tasks. This trend is also observed in the domain of code understanding and generation. Numerous Code LLMs [11; 12; 13; 14; 15; 16; 17; 18] have been proposed to tackle the challenges associated with code-related tasks. These Code LLMs undergo pre-training using substantial amounts of code data, enabling them to excel in various code-related tasks, showcasing impressive performance. In contrast to most previous Code LLMs that primarily emphasize the pre-training process, there has been limited exploration of fine-grained instruction tuning in the Code domain. The introduction of instruction tuning initially aimed to enhance the generalization capabilities of LMs across different tasks [19; 20; 21; 22; 23; 24; 25]. OpenAI's InstructGPT [10], for instance, involved soliciting human annotators to provide explicit instructions to ensure alignment with users' intentions. Similarly, recent works such as Alpaca [26] employed the self-instruct [27] method, where ChatGPT generated the instruction data. Vicuna [28] utilized user-shared conversations collected from ShareGPT.com. WizardLM [29] introduced the _Evol-Instruct_ method, which involved evolving existing instruction data to generate more complex and diverse datasets. However, it is worth noting that all these approaches primarily focused on the general domain and lacked specific design considerations for the code domain. Motivated by the _Evol-Instruct_ method, this study aims to enhance the capabilities of the SOTA open-source Code LLM, StarCoder [11], by generating intricate code instruction data through code-specific _Evol-Instruct_. To achieve this, we have made several adaptations to the evolutionary prompt process tailored specifically for code-related tasks. These modifications include refining the evolutionary instructions, simplifying the form of evolutionary prompts, and incorporating code debugging and time-space complexity constraints. Initially, our method is applied to evolve the basic code instruction data, Code Alpaca [30]. Subsequently, we conduct fine-tuning of StarCoder using our newly created code instruction-following training set and obtain our _WizardCoder_. The experimental results obtained from four code generation benchmarks, namely HumanEval [31], HumanEval+ [32], MBPP [33], and DS-100 [34], demonstrate that our _WizardCoder_ outperforms all other open-source Code LLMs, achieving state-of-the-art (SOTA) performance. Specifically, we observe a substantial improvement in pass@1 scores, with an increase of +22.3 (57.3 vs. 35.0) in HumanEval and +8.2 (51.8 vs. 43.6) in MBPP. Remarkably, despite its much smaller size, our _WizardCoder_ even surpasses Anthropic's Claude and Google's Bard in terms of pass rates on HumanEval and HumanEval+. The contributions of this work can be summarized as follows: * We introduce _WizardCoder_, which enhances the performance of the open-source Code LLM, StarCoder, through the application of Code _Evol-Instruct_. * _WizardCoder_ surpasses all other open-source Code LLMs by a substantial margin in terms of code generation, including StarCoder, CodeGen, CodeGee, CodeT5+, InstructCodeT5+, StarCoder-GPTeacher, and Instruct-Codegen-16B. * _WizardCoder_ achieves superior results in code generation compared to the largest closed-source LLMs, such as Claude, Bard, PaLM, PaLM-2, and LaMDA, despite being considerably smaller in size. ## 2 Related Work Large Language Models.Recently, LLMs have demonstrated remarkable achievements across a broad spectrum of tasks. Prominent tech companies have made significant strides in developing highly proficient LLMs. These include OpenAI's GPT3&4 [1; 2], Google's PaLM [3; 4], and Bard3, DeepMind's Chinchilla [5], and Gopher [6], as well as Anthropic's Claude4. However, it is important to note that these models are closed-source and can only be accessed through specific APIs or may not be accessible at all. Footnote 3: [https://bard.google.com/](https://bard.google.com/) Footnote 4: [https://www.anthropic.com/index/introducing-claude](https://www.anthropic.com/index/introducing-claude) The AI community has witnessed the release of several open-source LLMs, where the model weights are made publicly available. EleutherAI has contributed GPT-NeoX-20B [35] and GPT-J-GB [36]. Google has released UL2-20B [37]. Tsinghua University has introduced GLM-130B [7]. Meta has released OPT [9] and LLaMA [8]. It is worth noting that while these open-source models have made valuable contributions, they generally do not exhibit the same level of performance as their closed-source counterparts. Large Language Models for Code.Recent studies have introduced a significant number of LLMs for code-related tasks to address the challenges of code understanding and generation. OpenAI has unveiled Codex [16] and Code-Davinci [38]. Google has proposed PaLM-Coder [3]. They perform outstandingly on the popular code completion benchmarks, like HumanEval [31] and MBPP [33]. However, these models are closed-source. On the other hand, there are several open-source Code LLMs available. Salesforce has introduced CodeGen [13], CodeT5 [17], and CodeT5+ [18]. Tsinghua University has contributed CodeGeeX [14], and the BigCode Project has developed StarCoder [11]. These models have demonstrated notable advancements in code-related tasks. However, when compared to the SOTA closed-source models, they still lag behind significantly. In contrast to the aforementioned models without instruction fine-tuning, our work demonstrates that further training Code LLMs with Code _Evol-Instruct_ can substantially enhance performance. Instruction Fine-Tuning.The primary objective of instruction fine-tuning in its early stages was to enhance the cross-task generalization capabilities of LMs. This was achieved by fine-tuning LMs with a substantial corpus of public NLP tasks. T5 [19] was among the first models to explore this approach, training on a multitude of supervised text-to-text tasks. Subsequent works such as FLAN [20], ExT5 [22], T0 [23], and UnifiedQA [25] further expanded the range of tasks to bolster the overall generalization ability of LMs. Notably, ZeroPrompt [24] and FLAN-T5 [21] pushed the envelope by incorporating thousands of tasks in their training pipelines. Across these studies, a consistent finding emerges: fine-tuning LMs with diverse NLP task instructions yields significant performance improvements when applied to new tasks. While fine-tuning LMs with diverse NLP tasks has shown promising results, it often falls short in aligning with the intentions of real-world users. OpenAI has pursued a different approach by soliciting human annotators to provide a large corpus of human instructions, encompassing diverse forms and a wide range of task types. Building upon this dataset, OpenAI trained its GPT3 [1] model to create InstructGPT [10], which better aligns with users' inputs. This line of development has even led to the impressive work known as ChatGPT. However, it is important to note that the dataset and model weights associated with these advancements are not publicly available. Alpaca [26] takes a different route by adopting the self-instruct method [27], leveraging ChatGPT to generate data for training. Vicuna [28] utilizes user-shared conversations collected from ShareGPT.com to train its models. WizardLM [29] introduces the _Evol-Instruct_ method, which involves evolving existing instruction data to generate more complex and diverse datasets. In contrast to these general instruction fine-tuning approaches, our _WizardCoder_ successfully applies the _Evol-Instruct_ method specifically in the domain of Code LLMs. ## 3 Approach In this section, we elaborate on the methodological details of _WizardCoder_. Following WizardLM, we apply the _Evol-Instruct_ method to evolve Code Alpaca generated using self-instruct and fine-tune the pre-trained Code LLM StarCoder with the evolved data. ### Evol-Instruct Prompts for Code Inspired by the Evol-Instruct [29] method proposed by WizardLM, this work also attempts to make code instructions more complex to enhance the fine-tuning effectiveness of code pre-trained large models. To adapt Evol-Instruct to the realm of code, we made the following modifications to the evolutionary prompt: 1. Streamlined the evolutionary instructions by removing deepening, complicating input, and In-Breadth Evolving. 2. Simplified the form of evolutionary prompts by unifying the evolutionary prompt template. 3. Addressing the specific characteristics of the code domain, we added two evolutionary instructions: code debugging and code time-space complexity constraints. The unified code evolutionary prompt template is as follows: Please increase the difficulty of the given programming test question a bit. You can increase the difficulty using, but not limited to, the following methods: {method} {question} Here, {question} represents the current code instruction awaiting evolution, and {method} is the type of evolution. The five types we used are listed as follows:Add new constraints and requirements to the original problem, adding approximately 10 additional words. Replace a commonly used requirement in the programming task with a less common and more specific one. If the original problem can be solved with only a few logical steps, please add more reasoning steps. Provide a piece of erroneous code as a reference to increase misdirection. Propose higher time or space complexity requirements, but please refrain from doing so frequently. ### Training _WizardCoder_ We employ the following procedure to train _WizardCoder_. Initially, we utilize StarCoder 15B [11] as the foundation and proceed to fine-tune it using the code instruction-following training set, which was evolved through _Evol-Instruct_. The prompt format for fine-tuning is outlined as follows: Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request. ### Instruction: {instruction} ### Response: To construct the training dataset, we initialized it with the 20K instruction-following dataset called Code Alpaca5. We iteratively employ the Evol-Instruct technique on this dataset consisting of 20,000 samples to produce evolved data. After each round of data evolution, we merge the evolved data from all previous rounds with the original dataset to finetune StarCoder and assess the pass@1 metric on HumanEval [31]. Once we observe a decline in the pass@1 metric, we will discontinue the usage of Evol-Instruct and choose the model with the highest pass@1 as the ultimate model. Footnote 5: [https://github.com/sahil280114/codealpaca](https://github.com/sahil280114/codealpaca) ## 4 Experiment This section begins by providing a comprehensive overview of the baseline models in our experiments. Subsequently, we present the performance of our models on four code generation benchmarks: HumanEval [31], HumanEval+ [32], MBPP [33], and DS-1000 [34]. ### Baselines Closed-Source Models.Multiple technology companies have successfully developed highly proficient LLMs while choosing not to publicly release them. These models are referred to as closed-source models. For our research, we incorporate a substantial number of these models as our baselines. Specifically, our baselines encompass the following: (i) OpenAI's GPT3.5&4 [2], Code-Davinci-002 [38], Code-Cushman-001 [38], and Codex [16]; (ii) Google's Bard, PaLM 2 [4], PaLM [3], and LaMDA [40]; (iii) Google DeepMind's AlphaCode [12]; and (iv) Anthropic's Claude. Open-Source Models.Several open-source LLMs have been made available to the AI community, although their performance generally lags behind the closed-source models a lot. As part of our research, we incorporate a significant number of these open-source models as our baselines. Our baselines encompass the following models: StarCoder [11], LLaMa [8], CodeGen [13],CodeGeeX [14], CodeT5+[18], and InCoder[15]. In addition, we also include several models with instructions fine-tuning, including StarCoder-GPTeacher,6 Instruct-Codegen-16B,7 Guanaco-65B,8 and Falcon-40B-Instruct.9 Footnote 6: [https://huggingface.co/GeorgiaTechResearchInstitute/starcoder-gpteacher-code-instruct](https://huggingface.co/GeorgiaTechResearchInstitute/starcoder-gpteacher-code-instruct) Footnote 7: [https://huggingface.co/sahil2801/instruct-codegen-16B](https://huggingface.co/sahil2801/instruct-codegen-16B) Footnote 8: [https://huggingface.co/TheBloke/guanaco-65B-HF](https://huggingface.co/TheBloke/guanaco-65B-HF) Footnote 9: [https://huggingface.co/tiuae/falcon-40b-instruct](https://huggingface.co/tiuae/falcon-40b-instruct) ### Implementation Details The StarCoder [11] serves as our basic foundation model. The evolved dataset consists of approximately 78k samples. To fine-tune the basic models, we employ specific configurations, including a batch size of 512, a sequence length of 2048, 200 fine-tuning steps, 30 warmup steps, a learning rate of 2e-5, a Cosine learning rate scheduler, and fp16 mixed precision. ### Evaluation on HumanEval, HumanEval+, and MBPP HumanEval [31], HumanEval+ [32] and MBPP [33] are extensively utilized benchmarks within the field of Code LLMs. These benchmarks encompass a vast collection of Python programming problems, employing test cases to validate the code generated by Code LLMs. HumanEval consists of 164 original programming problems, with an average of 9.6 test cases allocated to each problem. To ensure a thorough assessment of the functional correctness of LLM-synthesized code, HumanEval+ extends the number of test cases significantly, averaging at 774.8 test cases per problem. On the other hand, MBPP offers a set of 500 test programming problems, accompanied by three automated test cases per problem. The prompt format for these tasks is as follows: Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request. ### Instruction: Create a Python script for this problem: {Question} ### Response: Figure 1: The percentage of pass rates on the HumanEval (164 problems) with a single attempt. All baseline scores are retrieved from the LLM-Humaneval-Benchmarks [39]. Our _WizardCoder_ generates an answer with greedy decoding. Comparing with the Closed-Source Models.The SOTA LLMs for code generation, such as GPT4, Claude, and Bard, are predominantly closed-source. Acquiring access to the APIs of these models proves challenging. In this study, we adopt an alternative approach by retrieving the scores for HumanEval and HumanEval+ from the LLM-Humaneval-Benchmarks [39]. Notably, all the mentioned models generate code solutions for each problem utilizing a single attempt, and the resulting pass rate percentage is reported. To maintain consistency, we employ the same experimental setup by generating answers using greedy decoding and evaluate our _WizanLoder_ using the provided evaluation codes. By adhering to these standardized procedures, we aim to ensure fair and comparable evaluations of our model against existing benchmarks. As depicted in Figure 1, our _WizardCoder_ attains the third position in this benchmark, surpassing Claude-Plus (59.8 vs. 53.0) and Bard (59.8 vs. 44.5). Notably, our model exhibits a substantially smaller size compared to these models. Furthermore, our _WizandCoder_ demonstrates a remarkable superiority over other open-source LLMs that undergo instruction fine-tuning, showcasing a significant performance margin. Comparing with the Open-Source Models.In Table 1, we conduct a comprehensive comparison of our _WizardCoder_ with other open-source models on the HumanEval and MBPP benchmarks. In contrast to the results presented in Figure 1, we adhere to the approach outlined in previous studies [31] by generating n samples for each problem to estimate the pass@1 score. The findings presented in Table 1 clearly demonstrate that our _WizardCoder_ exhibits a substantial performance advantage over all the open-source models. From the experimental results in Figure 1 and Table 1, we have the following conclusions: 1. _WizardCoder_ outperforms the largest closed-source LLMs, including Claude, Bard, PaLM, PaLM-2, and LaMDA, despite being significantly smaller. \begin{table} \begin{tabular}{l c c c} \hline \hline **Model** & **Params** & **HumanEval** & **MBPP** \\ \hline \multicolumn{4}{c}{Closed-source models} \\ \hline LaMDA [40] & 137B & 14.0 & - \\ AlphaCode [12] & 1.1B & 17.1 & - \\ PaLM [3] & 540B & 26.2 & 36.8 \\ PaLM-Coder [3] & 540B & 36.0 & 47.0 \\ PaLM 2-S [4] & - & 37.6 & 50.0 \\ Codex [16] & 2.5B & 21.4 & - \\ Codex [16] & 12B & 28.8 & - \\ Code-Cushman-001 [38] & - & 33.5 & 45.9 \\ Code-Davinci-002 [38] & - & 47.0 & 58.1 \\ GPT-3.5 [2] & - & 48.1 & - \\ GPT-4 [2] & - & 67.0 & - \\ \hline \multicolumn{4}{c}{Open-source models} \\ \hline LLaMa [8] & 33B & 21.7 & 30.2 \\ LLaMa [8] & 65B & 23.7 & 37.7 \\ CodeGen-Multi [13] & 16B & 18.3 & 20.9 \\ CodeGen-Mono [13] & 16B & 29.3 & 35.3 \\ CodeGeeX [14] & 13B & 22.9 & 24.4 \\ StarCoder [11] & 15B & 33.6 & 43.6\({}^{*}\) \\ CodeT5+ [18] & 16B & 30.9 & - \\ InstructCodeT5+ [18] & 16B & 35.0 & - \\ \hline _WizardCoder_ & 15B & **57.3** (+22.3) & **51.8** (+8.2) \\ \hline \hline \end{tabular} \end{table} Table 1: Results of pass@1(%) on HumanEval and MBPP. Most scores are retrieved from the papers of StarCoder [11] and CodeT5+ [18]. We follow the previous works [31] to generate n samples to estimate the pass@1 score with the same set of hyper-parameters: temperate=0.2, and top_p=0.95. *: we evaluate this model by ourselves. 2. _WizardCoder_ outperforms all the open-source Code LLMs by a large margin (+22.3 on HumanEval), including StarCoder, CodeGen, CodeGee, and CodeT5+. 3. _WizardCoder_ significantly outperforms all the open-source Code LLMs with instructions fine-tuning, including InstructCodeT5+, StarCoder-GPTeacher, and Instruct-Codegen-16B. ### Evaluation on DS-1000 The DS-1000 benchmark [34] comprises 1,000 distinct data science workflows spanning seven libraries. It assesses the performance of code generations against test cases and supports two evaluation modes: completion and insertion. In our experiments, we only report insertion scores for models that support. The DS-1000 benchmark further classifies problems based on the libraries employed, including Matplotlib (plt), NumPy (np), Pandas (pd), SciPy (scp), Scikit-Learn (sk), PyTorch (py), and TensorFlow (tf). We follow the same prompt format as StarCoder. In Table 2, we present pass@1 (n=40) results for each library, along with an overall score. Based on these results, our conclusion is that _WizardCoder_ demonstrates a significant superiority over all other models when tackling data science problems on the DS-1000 benchmark. This observation holds true across nearly all data science libraries. ### Ablation Study Figure 2 presents an ablation study investigating the impact of the number of data evolution rounds. The first round of evolved data contains 38k samples. The second round contains 58k. The third round contains 78k. The fourth round contains 98k. For consistency, all models undergo fine-tuning with 200 steps. The results reveal that the highest pass@1 score on humaneval is achieved after three rounds of data evolution. Based on this observation, we select the data that evolved during the third round as the ultimate dataset. \begin{table} \begin{tabular}{l c c c c c c c c c} \hline \hline **Format** & **Model** & **plt** & **np** & **pd** & **py** & **scp** & **sk** & **tf** & **All** \\ \hline & \# of problems: & 155 & 220 & 291 & 68 & 106 & 115 & 45 & 1,000 \\ \hline Completion & InCoder-6B & 28.3 & 4.4 & 3.1 & 4.4 & 2.8 & 2.8 & 3.8 & 7.4 \\ Completion & CodeGen-mono & 31.7 & 10.9 & 3.4 & 7.0 & 9.0 & 10.8 & 15.2 & 11.7 \\ Completion & Code-Cushman-001 & 40.7 & 21.8 & 7.9 & 12.4 & 11.3 & 18.0 & 12.2 & 18.1 \\ Completion & StarCoder & 51.7 & 29.7 & 11.4 & 21.4 & 20.2 & **29.5** & 24.5 & 26.0 \\ Completion & _WizardCoder_ & **55.2** & **33.6** & **16.7** & **26.2** & **24.2** & 24.9 & **26.7** & **29.2** \\ \hline Insertion & InCoder-6B & 28.3 & 4.6 & 2.9 & 4.4 & 2.8 & 3.1 & 7.8 & 7.5 \\ Insertion & StarCoder & 51.7 & 30.8 & 10.3 & 21.0 & 20.2 & 27.4 & 20.0 & 25.4 \\ Insertion & _WizardCoder_ & **55.2** & **35.1** & **20.4** & **30.4** & **28.9** & **32.3** & **37.8** & **32.8** \\ \hline \hline \end{tabular} \end{table} Table 2: Performance of _WizardCoder_ and baseline models on DS-1000. All models are evaluated with the same set of hyper-parameters: temperature=0.2, top_p=0.5, max_length=1024. Scores are average pass@1 accuracy over 40 samples. Matplotlib (plt) task does not have the right context, so insertion and completion scores are identical. Figure 2: Ablation study on the number of data evolution rounds. ### Examples Table 3 showcases examples of interactions with our _WizardCoder_. The examples demonstrate that our model consistently generates accurate responses accompanied by clear explanations. ## 5 Conclusion and Future Work This paper introduces _WizardCoder_, a Code _Evol-Instruct_ fine-tuned Code LLM. The experimental results demonstrate that _WizardCoder_ achieves SOTA performance surpassing all existing open-source Code LLMs on four widely recognized code generation benchmarks: HumanEval, HumanEval+, MBPP, and DS-1000. Furthermore, _WizardCoder_ exhibits superior performance compared to the largest closed LLMs, including Anthropic's Claude and Google's Bard. Future Work.Although our _WizardCoder_ demonstrates impressive coding performance, as depicted in Figure 1, our model still falls significantly behind the SOTA LLM, GPT4. Therefore, future work will prioritize the enhancement of the Code _Evol-Instruct_ method to further augment the performance of our model. Broader Impact.Similar to the other LLMs, our _WizardCoder_ could also generate unethical, harmful, or misleading information. Therefore, future research to address the ethical and societal implications is needed. \begin{table} \begin{tabular}{l|l} \hline \hline **Instruction** & **Response** \\ \hline \hline \multirow{7}{*}{Write a Python function to tell me what the date is today.} & Here’s a Python function that tells you what the date is today: \\ & import:dattime \\ & dart get_data(): \\ & now = datetime.datetime.now() \\ & return now.ntrfftime(’L1, L1 Kd., T1’) \\ & print(get_data()): \\ \cline{1-1} & This function uses the datetime module to get the current date and time. It then uses the strftime() method to format the date as a string in the format "Day, Month Date, Year". Finally, it returns the formatted string. \\ \cline{1-1} & You can call this function from your Python code to get the current date and time. \\ \hline \multirow{7}{*}{This code takes in a function f, the lower bound a, the upper bound b, and the number of subintervals n. It then calculates the width of each subinterval h, and uses the trapezoidal rule to approximate the integral of f between a and b. The result is returned as a float. \\ \cline{1-1} & To use this function, you would call it with the function you want to integrate, the lower bound, the upper bound, and the number of subintervals you want to use. For example: \\ \cline{1-1} & dart f(x): \\ \cline{1-1} & return **2 \\ \cline{1-1} & result = integrate(f. 0, 1, 1000) \\ \cline{1-1} & This would integrate the function f(x) = x**2 between 0 and 1 using 1000 subintervals, and print the result. \\ \hline \hline \multirow{7}{*}{List the prime numbers between 20 and 30 with Java.} & \\ \cline{1-1} & & \\ \cline{1-1} & & \\ \cline{1-1} & & \\ \cline{1-1} & & \\ \cline{1-1} & & \\ \cline{1-1} & & \\ \cline{1-1} & & \\ \cline{1-1} & & \\ \hline \hline \end{tabular} \end{table} Table 3: Examples of interaction with our _WizardCoder_. ## References * [1] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In Hugo Larochelle, Marc'Aurelio Ranzato, Raia Hadsell, Maria-Florina Balcan, and Hsuan-Tien Lin, editors, _Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual_, 2020. * [2] OpenAI. GPT-4 technical report. _CoRR_, abs/2303.08774, 2023. * [3] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, Parker Schuh, Kensen Shi, Sasha Tsvyashchenko, Joshua Maynez, Abhishek Rao, Parker Barnes, Yi Tay, Noam Shazeer, Vinodkumar Prabhakaran, Emily Reif, Nan Du, Ben Hutchinson, Reiner Pope, James Bradbury, Jacob Austin, Michael Isard, Guy Guor-Ari, Pengcheng Yin, Toju Duke, Anselm Levskaya, Sanjay Ghemawat, Sunjna Dev, Henry Michalewski, Xavier Garcia, Vedant Misra, Kevin Robinson, Liam Fedus, Denmy Zhou, Daphne Ippolito, David Luan, Hyeontaek Lim, Barret Zoph, Alexander Spiridonov, Ryan Sepassi, David Dohan, Shivani Agrawal, Mark Omernick, Andrew M. Dai, Thanumalayan Sankaranarayana Pillai, Marie Pellat, Aitor Lewkowycz, Eric Moreira, Rewon Child, Oleksandr Polozov, Katherine Lee, Zongwei Zhou, Xuezhi Wang, Brennan Saeta, Mark Diaz, Orhan Firat, Michele Catasta, Jason Wei, Kathy Meier-Hellstern, Douglas Eck, Jeff Dean, Slav Petrov, and Noah Fiedel. Palm: Scaling language modeling with pathways. _CoRR_, abs/2204.02311, 2022. * [4] Rohan Anil, Andrew M. Dai, Orhan Firat, Melvin Johnson, Dmitry Lepikhin, Alexandre Passos, Siamak Shakeri, Emanuel Taropa, Paige Bailey, Zhifeng Chen, Eric Chu, Jonathan H. Clark, Laurent El Shafey, Yanping Huang, Kathy Meier-Hellstern, Gaurav Mishra, Erica Moreira, Mark Omernick, Kevin Robinson, Sebastian Ruder, Yi Tay, Kefan Xiao, Yuanzhong Xu, Yujing Zhang, Gustavo Hernandez Abrego, Junwhan Ahn, Jacob Austin, Paul Barham, Jan A. Botha, James Bradbury, Siddhartha Brahma, Kevin Brooks, Michele Catasta, Yong Cheng, Colin Cherry, Christopher A. Choquette-Choo, Aakanksha Chowdhery, Clement Crepy, Shachi Dave, Mostafa Dehghani, Sunipa Dev, Jacob Devlin, Mark Diaz, Nan Du, Ethan Dyer, Vladimir Feinberg, Fangxiaoyu Feng, Vlad Fienberg, Markus Freitag, Xavier Garcia, Sebastian Gehrmann, Lucas Gonzalez, and et al. Palm 2 technical report. _CoRR_, abs/2305.10403, 2023. * [5] Jordan Hoffmann, Sebastian Borgeaud, Arthur Mensch, Elena Buchatskaya, Trevor Cai, Eliza Rutherford, Diego de Las Casas, Lisa Anne Hendricks, Johannes Welbl, Aidan Clark, Tom Hennigan, Eric Noland, Katie Millican, George van den Driessche, Bogdan Damoc, Aurelia Guy, Simon Osindero, Karen Simonyan, Erich Elsen, Jack W. Rae, Oriol Vinyals, and Laurent Sifre. Training compute-optimal large language models. _CoRR_, abs/2203.15556, 2022. * [6] Jack W. Rae, Sebastian Borgeaud, Trevor Cai, Katie Millican, Jordan Hoffmann, H. Francis Song, John Aslanides, Sarah Henderson, Roman Ring, Susannah Young, Eliza Rutherford, Tom Hennigan, Jacob Menick, Albin Casser, Richard Powell, George van den Driessche, Lisa Anne Hendricks, Maribeth Rauh, Po-Sen Huang, Amelia Glasee, Johannes Welbl, Sumanth Dathathathri, Saffron Huang, Jonathan Desato, John Mellor, Irina Higgins, Antonia Creswell, Nat McAleese, Amy Wu, Erich Elsen, Siddhant M. Jayakumar, Elena Buchatskaya, David Budden, Esme Sutherland, Karen Simonyan, Michela Paganini, Laurent Sifre, Lena Martens, Xiang Lorraine Li, Adhiguna Kuncoro, Aida Nematzadeh, Elena Gribovskaya, Domenic Donato, Angeliki Lazaridou, Arthur Mensch, Jean-Baptiste Lepsiau, Maria Tsimpoukelli, Nikola Grigorev, Doug Fritz, Thibault Sotititauv, Manus Palayarskas, Toby Pohlen, Zhitao Gong, Daniel Toyama, Cyprien de Masson d'Autume, Yujia Li, Tayfun Terzi, Vladimir Mikulik, Igor Babuschkin, Aidan Clark, Diego de Las Casas, Aurelia Guy, Chris Jones, James Bradbury, Matthew J. Johnson, Blake A. Hechtman, Laura Weidinger, Iason Gabriel, William Isaac, Edward Lockhart, Simon Osindero, Laura Rimell, Chris Dyer, Oriol Vinyals, Kareem Ayoub, Jeff Stanway, Lorrayne Bennett, Demis Hassabis, Koray Kavukcuoglu, and Geoffrey Irving. Scaling language models: Methods, analysis & insights from training gopher. _CoRR_, abs/2112.11446, 2021. * [7] Aohan Zeng, Xiao Liu, Zhengxiao Du, Zihan Wang, Hanyu Lai, Ming Ding, Zhuoyi Yang, Yifan Xu, Wendi Zheng, Xiao Xia, Weng Lam Tam, Zixuan Ma, Yufei Xue, Jidong Zhai, Wenguang Chen, Peng Zhang, Yuxiao Dong, and Jie Tang. GLM-130B: an open bilingual pre-trained model. _CoRR_, abs/2210.02414, 2022. * [8] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, ArmandJoulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation language models. _CoRR_, abs/2302.13971, 2023. * [9] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona T. Diab, Xian Li, Xi Victoria Lin, Todor Mihaylov, Myle Ott, Sam Shleifer, Kurt Shuster, Daniel Simig, Punit Singh Koura, Anjali Sridhar, Tianlu Wang, and Luke Zettlemoyer. OPT: open pre-trained transformer language models. _CoRR_, abs/2205.01068, 2022. * [10] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul F. Christiano, Jan Leike, and Ryan Lowe. Training language models to follow instructions with human feedback. In _NeurIPS_, 2022. * [11] Raymond Li, Loubna Ben Allal, Yangtian Zi, Niklas Muennighoff, Denis Kocetkov, Chenghao Mou, Marc Marone, Christopher Akiki, Jia Li, Jenny Chim, et al. Starcoder: may the source be with you! _arXiv preprint arXiv:2305.06161_, 2023. * [12] Yujia Li, David H. Choi, Junyoung Chung, Nate Kushman, Julian Schrittwieser, Remi Leblond, Tom Eccles, James Keeling, Felix Gimeno, Agustin Dal Lago, Thomas Hubert, Peter Choy, Cyprien de Masson d'Autume, Igor Babuschkin, Xinyun Chen, Po-Sen Huang, Johannes Welbl, Sven Gowal, Alexey Cherepanov, James Molloy, Daniel J. Mankowitz, Esme Sutherland Robson, Pushmeet Kohli, Nando de Freitas, Koray Kavukcuoglu, and Oriol Vinyals. Competition-level code generation with alphacode. _CoRR_, abs/2203.07814, 2022. * [13] Erik Nijkamp, Bo Pang, Hiroaki Hayashi, Lifu Tu, Huan Wang, Yingbo Zhou, Silvio Savarese, and Caiming Xiong. Codegen: An open large language model for code with multi-turn program synthesis. In _The Eleventh International Conference on Learning Representations_, 2023. * [14] Qinkai Zheng, Xiao Xia, Xu Zou, Yuxiao Dong, Shan Wang, Yufei Xue, Zihan Wang, Lei Shen, Andi Wang, Yang Li, Teng Su, Zhilin Yang, and Jie Tang. Codegeex: A pre-trained model for code generation with multilingual evaluations on humaneval-x. _CoRR_, abs/2303.17568, 2023. * [15] Daniel Fried, Armen Aghajanyan, Jessy Lin, Sida Wang, Eric Wallace, Freda Shi, Ruiqi Zhong, Wen-tau Yih, Luke Zettlemoyer, and Mike Lewis. Incoder: A generative model for code infilling and synthesis. _CoRR_, abs/2204.05999, 2022. * [16] Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harrison Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, Alex Ray, Raul Puri, Gretchen Krueger, Michael Petrov, Heidy Khlaaf, Girish Sastry, Pamela Mishkin, Brooke Chan, Scott Gray, Nick Ryder, Mikhail Pavlov, Alethea Power, Lukasz Kaiser, Mohammad Bavarian, Clemens Winter, Philippe Tillet, Felipe Petroski Such, Dave Cummings, Matthias Plappert, Fotos Chantzis, Elizabeth Barnes, Ariel Herbert-Voss, William Hebgen Guss, Alex Nichol, Alex Piano, Nikolas Tezak, Jie Tang, Igor Babuschkin, Suchir Balaji, Shantanu Jain, William Saunders, Christopher Hesse, Andrew N. Carr, Jan Leike, Joshua Achiam, Vedant Misra, Evan Morikawa, Alec Radford, Matthew Knight, Miles Brundage, Mira Murati, Katie Mayer, Peter Welinder, Bob McGrew, Dario Amodei, Sam McCandish, Ilya Sutskever, and Wojciech Zaremba. Evaluating large language models trained on code. _CoRR_, abs/2107.03374, 2021. * [17] Yue Wang, Weishi Wang, Shafiq R. Joty, and Steven C. H. Hoi. Coded5: Identifier-aware unified pre-trained encoder-decoder models for code understanding and generation. In Marie-Francine Moens, Xuuning Huang, Lucia Specia, and Scott Wen-tau Yih, editors, _Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, EMNLP 2021, Virtual Event / Punta Cana, Dominican Republic, 7-11 November, 2021_, pages 8696-8708. Association for Computational Linguistics, 2021. * [18] Yue Wang, Hung Le, Akhilesh Deepak Gotmare, Nghi D. Q. Bui, Junnan Li, and Steven C. H. Hoi. Coded5+: Open code large language models for code understanding and generation. _CoRR_, abs/2305.07922, 2023. * [19] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. _J. Mach. Learn. Res._, 21:140:1-140:67, 2020. * [20] Jason Wei, Maarten Bosma, Vincent Y. Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M. Dai, and Quoc V. Le. Finetuned language models are zero-shot learners. In _The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022_. OpenReview.net, 2022. * [21] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirca Suzgun, Xinyun Chen, Aakanska Chowdrey, Sharan Narang, Gaurav Mishra, Adams Yu, Vincent Y. Zhao, Yanping Huang, Andrew M. Dai, Hongkun Yu, Slav Petrov, Ed H. Chi, Jeff Dean, Jacob Devlin, Adam Roberts, Denny Zhou, Quoc V. Le, and Jason Wei. Scaling instruction-finetuned language models. _CoRR_, abs/2210.11416, 2022. * [22] Vamsi Aribandi, Yi Tay, Tal Schuster, Jinfeng Rao, Huaixiu Steven Zheng, Sanket Vaibhav Mehta, Honglei Zhuang, Vinh Q. Tran, Dara Bahri, Jianmo Ni, Jia Prakash Gupta, Kai Hui, Sebastian Ruder, and Donald Metzler. Ext5: Towards extreme multi-task scaling for transfer learning. In _The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022_. OpenReview.net, 2022. * [23] Victor Sanh, Albert Webson, Colin Raffel, Stephen H. Bach, Lintang Sutawika, Zaid Alyafeai, Antoine Chaffin, Arnaud Stiegler, Arun Raja, Manan Dey, M Saiful Bari, Canwen Xu, Urmish Thakker, Shanya Sharma Sharma, Eliza Szczechla, Taewoon Kim, Gunjan Chhablani, Nihal V. Nayak, Debajoti Datta, Jonathan Chang, Mike Tian-Jian Jiang, Han Wang, Matteo Manica, Sheng Shen, Zheng Xin Yong, Harshif Pandey, Rachel Bawden, Thomas Wang, Trishala Neeraj, Jose Rozen, Abbeesh Sharma, Andrea Santilli, Thibault Fevry, Jason Alan Fries, Ryan Techan, Teven Le Scao, Stella Biderman, Leo Gao, Thomas Wolf, and Alexander M. Rush. Multitask prompted training enables zero-shot task generalization. In _The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022_. OpenReview.net, 2022. * [24] Hanwei Xu, Yujun Chen, Yulun Du, Nan Shao, Yanggang Wang, Haiyu Li, and Zhilin Yang. Zeroprompt: Scaling prompt-based pretraining to 1, 000 tasks improves zero-shot generalization. In Yoav Goldberg, Zornitsa Kozareva, and Yue Zhang, editors, _Findings of the Association for Computational Linguistics: EMNLP 2022, Abu Dhabi, United Arab Emirates, December 7-11, 2022_, pages 4235-4252. Association for Computational Linguistics, 2022. * [25] Daniel Khashabi, Sewon Min, Tushar Khot, Ashish Sabharwal, Oyvind Tafjord, Peter Clark, and Hannaneh Hajishirzi. Unifiedqa: Crossing format boundaries with a single QA system. In Trevor Cohn, Yulan He, and Yang Liu, editors, _Findings of the Association for Computational Linguistics: EMNLP 2020, Online Event, 16-20 November 2020_, volume EMNLP 2020 of _Findings of ACL_, pages 1896-1907. Association for Computational Linguistics, 2020. * [26] Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B. Hashimoto. Stanford alpaca: An instruction-following llama model. [https://github.com/tatsu-lab/stanford_alpaca](https://github.com/tatsu-lab/stanford_alpaca), 2023. * [27] Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A Smith, Daniel Khashabi, and Hannaneh Hajishirzi. Self-instruct: Aligning language model with self generated instructions. _arXiv preprint arXiv:2212.10560_, 2022. * [28] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with 90%* chatgtp quality, March 2023. * [29] Can Xu, Qingfeng Sun, Kai Zheng, Xiubo Geng, Pu Zhao, Jiazhan Feng, Chongyang Tao, and Daxin Jiang. Wizardlm: Empowering large language models to follow complex instructions. _arXiv preprint arXiv:2304.12244_, 2023. * [30] Sahil Chaudhary. Code alpaca: An instruction-following llama model for code generation. [https://github.com/sahil280114/codealpaca](https://github.com/sahil280114/codealpaca), 2023. * [31] Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harrison Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, Alex Ray, Raul Puri, Gretchen Krueger, Michael Petrov, Heidy Khlaaf, Girish Sastry, Pamela Mishkin, Brooke Chan, Scott Gray, Nick Ryder, Mikhail Pavlov, Alethea Power, Lukasz Kaiser, Mohammad Bavarian, Clemens Winter, Philippe Tillet, Felipe Petroski Such, Dave Cummings, Matthias Plappert, Fotos Chantzis, Elizabeth Barnes, Ariel Herber-Voss, William Hebgen Guss, Alex Nichol, Alex Paino, Nikolas Tezak, Jie Tang, Igor Babuschkin, Suchir Balaji, Shantanu Jain, William Saunders, Christopher Hesse, Andrew N. Carr, Jan Leike, Joshua Achiam, Vedant Misra, Evan Morikawa, Alec Radford, Matthew Knight, Miles Brundage, Mira Murati, Katie Mayer, Peter Welinder, Bob McGrew, Dario Amodei, Sam McCandlish, Ilya Sutskever, and Wojciech Zaremba. Evaluating large language models trained on code. _CoRR_, abs/2107.03374, 2021. * [32] Jiawei Liu, Chunqiu Steven Xia, Yuyao Wang, and Lingming Zhang. Is your code generated by chatgpt really correct? rigorous evaluation of large language models for code generation. _CoRR_, abs/2305.01210, 2023. * [33] Jacob Austin, Augustus Odena, Maxwell I. Nye, Maarten Bosma, Henryk Michalewski, David Dohan, Ellen Jiang, Carrie J. Cai, Michael Terry, Quoc V. Le, and Charles Sutton. Program synthesis with large language models. _CoRR_, abs/2108.07732, 2021. * [34] Yuhang Lai, Chengxi Li, Yiming Wang, Tianyi Zhang, Ruiqi Zhong, Luke Zettlemoyer, Scott Wen-tau Yih, Daniel Fried, Sida I. Wang, and Tao Yu. DS-1000: A natural and reliable benchmark for data science code generation. _CoRR_, abs/2211.11501, 2022. * [35] Sid Black, Stella Biderman, Eric Hallahan, Quentin Anthony, Leo Gao, Laurence Golding, Horace He, Connor Leahy, Kyle McDonell, Jason Phang, Michael Pieler, USVSN Sai Prashanth, Shivanshu Purohit, Laria Reynolds, Jonathan Tow, Ben Wang, and Samuel Weinbach. Gpt-neox-20b: An open-source autoregressive language model. _CoRR_, abs/2204.06745, 2022. * [36] Ben Wang and Aran Komatsuzaki. GPT-J-6B: A 6 Billion Parameter Autoregressive Language Model. [https://github.com/kingoflolz/mesh-transformer-jax](https://github.com/kingoflolz/mesh-transformer-jax), May 2021. * [37] Yi Tay, Mostafa Dehghani, Vinh Q. Tran, Xavier Garcia, Dara Bahri, Tal Schuster, Huaixiu Steven Zheng, Neil Houlsby, and Donald Metzler. Unifying language learning paradigms. _CoRR_, abs/2205.05131, 2022. * [38] Microsoft. Azure openai service models. [https://learn.microsoft.com/en-us/azure/cognitive-services/openai/concepts/models](https://learn.microsoft.com/en-us/azure/cognitive-services/openai/concepts/models), 2023. * [39] Lim humaneval benchmarks. [https://github.com/my-other-github-account/llm-humaneval-benchmarks](https://github.com/my-other-github-account/llm-humaneval-benchmarks), 2023. * [40] Romal Thoppilan, Daniel De Freitas, Jamie Hall, Noam Shazeer, Apoorv Kulshreshtha, Heng-Tze Cheng, Alicia Jin, Taylor Bos, Leslie Baker, Yu Du, YaGuang Li, Hongrae Lee, Huaixiu Steven Zheng, Amin Ghafouri, Marcelo Menegal, Yanping Huang, Maxim Krikun, Dmitry Lepikhin, James Qin, Dehao Chen, Yuanzhong Xu, Zhifeng Chen, Adam Roberts, Maarten Bosma, Yanqi Zhou, Chung-Ching Chang, Igor Krivokon, Will Rusch, Marc Pickett, Kathleen S. Meier-Hellstern, Meredith Ringel Morris, Tulse Doshi, Reneilito Delos Santos, Toqju Duke, Johnny Soraker, Ben Zevenbergen, Vinodkumar Prabhakaran, Mark Diaz, Ben Hutchinson, Kristen Olson, Alejandra Molina, Erin Hoffman-John, Josh Lee, Lora Aroyo, Ravi Rajakumar, Alena Butryna, Matthew Lamm, Viktoriya Kuzmina, Joe Fenton, Aaron Cohen, Rachel Bernstein, Ray Kurzweil, Blaise Aguera-Arcas, Claire Cui, Marian Croak, Ed H. Chi, and Quoc Le. Lamda: Language models for dialog applications. _CoRR_, abs/2201.08239, 2022.

"# A Tale of Tails: Model Collapse as a Change of Scaling Laws\n\nElvis Dohmatob\n\nYunzhen Feng\n\n(...TRUNCATED)

"# Self-rag: Learning to Retrieve, Generate, and Critique through Self-Reflection\n\n Akari Asi\\({}(...TRUNCATED)

"[MISSING_PAGE_FAIL:1]\n\nstruction data generated by real human users, OpenAI's LLMs (e.g., Instruc(...TRUNCATED)

"# Fine Tuning vs. Retrieval Augmented Generation for Less Popular Knowledge\n\n Heydar Soudani\n\nR(...TRUNCATED)

"# Do Large Language Models Lantly Perform Multi-Hop Reasoning?\n\nSohee Yang1,2 Elena Gribovskaya1 (...TRUNCATED)

"# Quiet-STaR: Language Models Can Teach Themselves to Think Before Speaking\n\n Eric Zelikman\n\nSt(...TRUNCATED)

"# Nemotron-4 15B Technical Report\n\nJupinder Parmar Shrimai Prabhumoye Joseph Jennings Mostofa Pat(...TRUNCATED)

"# Language Model Inversion\n\nJohn X. Morris, Wenting Zhao, Justin T. Chiu, Vitaly Shmatikov, Alexa(...TRUNCATED)