ChemBench-Collection
Collection
Datasets, Spaces and Results related to ChemBench
•
3 items
•
Updated
•
2
model_id
string | name
string | number_params
int64 | description
string | date_published
string | paper_link
string | code_link
string | is_api_endpoint
bool | nr_of_tokens
int64 | architecture
string | is_open_weights
bool | is_open_dataset
bool | is_mixture_of_experts
bool | model_alignment
string | reinforcement_learning_from_human_feedback
bool | domain_specific_pretraining
bool | domain_specific_finetuning
bool | tool_use
bool | tool_type
string | temperature
float64 | epochs
int64 | reasoning_model
bool | reasoning_type
string | overall_score
float64 | Analytical Chemistry
float64 | Chemical Preference
float64 | General Chemistry
float64 | Inorganic Chemistry
float64 | Materials Science
float64 | Organic Chemistry
float64 | Physical Chemistry
float64 | Technical Chemistry
float64 | Toxicity and Safety
float64 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
mistral-large-2-123b | Mistral-Large-2 | 123,000,000,000 | Mistral Large 2 has a 128k context window and supports dozens of languages and along with 80+ coding languages. Mistral Large 2 is designed for single-node inference with long-context applications in mind. | 2024-07-24 | null | https://huggingface.co/mistralai/Mistral-Large-Instruct-2407 | true | null | DecoderOnly | true | false | false | null | null | false | false | false | null | 0 | null | false | null | 0.569943 | 0.480263 | 0.546454 | 0.785235 | 0.793478 | 0.666667 | 0.732558 | 0.690909 | 0.675 | 0.395556 |
llama3.1-70b-instruct | Llama-3.1-70B-Instruct | 70,000,000,000 | The Meta Llama 3.1 collection of multilingual large language models (LLMs) is a collection of pretrained and instruction tuned generative models in 8B, 70B and 405B sizes (text in/text out). The Llama 3.1 instruction tuned text only models (8B, 70B, 405B) are optimized for multilingual dialogue use cases and outperform many of the available open source and closed chat models on common industry benchmarks. | 2024-07-23 | https://arxiv.org/abs/2407.21783 | https://github.com/meta-llama/llama3 | true | 15,000,000,000,000 | DecoderOnly | true | false | false | DPO | true | false | false | false | null | 0 | null | false | null | 0.533716 | 0.407895 | 0.519481 | 0.691275 | 0.771739 | 0.666667 | 0.662791 | 0.642424 | 0.65 | 0.383704 |
claude3.5 | Claude-3.5 (Sonnet) | null | Claude models are general purpose large language models. They use a transformer architecture and are trained via unsupervised learning, RLHF, and Constitutional AI (including both a supervised and Reinforcement Learning (RL) phase). Claude 3.5 was developed by Anthropic. | 2024-06-20 | null | null | true | null | null | false | false | null | null | true | false | false | false | null | 0 | null | false | null | 0.625538 | 0.565789 | 0.584416 | 0.825503 | 0.836957 | 0.714286 | 0.825581 | 0.769697 | 0.85 | 0.44 |
mixtral-8x7b-instruct-T-one | Mixtral-8x7b-Instruct (Temperature 1.0) | 47,000,000,000 | Mixtral is a sparse mixture-of-experts network. It is a decoder-only model where the feedforward block picks from a set of 8 distinct groups of parameters. At every layer, for every token, a router network chooses two of these groups (the “experts”) to process the token and combine their output additively. | 2023-12-11 | https://arxiv.org/abs/2401.04088 | https://huggingface.co/mistralai/Mixtral-8x7B-v0.1 | true | null | DecoderOnly | true | false | true | DPO | false | false | false | false | null | 1 | null | false | null | 0.418221 | 0.276316 | 0.522478 | 0.449664 | 0.51087 | 0.404762 | 0.472093 | 0.345455 | 0.325 | 0.266667 |
command-r+ | Command-R+ | 104,000,000,000 | Cohere Command R is a family of highly scalable language models that balance high performance with strong accuracy. Command-R models were released by Cohere. | 2024-04-04 | null | https://huggingface.co/CohereForAI/c4ai-command-r-plus | true | null | null | false | false | null | null | true | false | false | false | null | 0 | null | false | null | 0.447633 | 0.342105 | 0.513487 | 0.496644 | 0.521739 | 0.464286 | 0.551163 | 0.327273 | 0.5 | 0.311111 |
gpt-4o-react | GPT-4o React | null | GPT-4o is OpenAI's third major iteration of their popular large multimodal model, GPT-4, which expands on the capabilities of GPT-4 with Vision. | 2024-05-13 | null | null | true | null | DecoderOnly | false | false | null | null | true | false | false | true | ArXiV, Web search, Wikipedia, Wolfram alpha calculator, SMILES to IUPAC name and IUPAC name to SMILES converters | 0 | null | false | null | 0.50538 | 0.467105 | 0.420579 | 0.758389 | 0.728261 | 0.559524 | 0.718605 | 0.6 | 0.725 | 0.374815 |
llama3.1-405b-instruct | Llama-3.1-405B-Instruct | 405,000,000,000 | The Meta Llama 3.1 collection of multilingual large language models (LLMs) is a collection of pretrained and instruction tuned generative models in 8B, 70B and 405B sizes (text in/text out). The Llama 3.1 instruction tuned text only models (8B, 70B, 405B) are optimized for multilingual dialogue use cases and outperform many of the available open source and closed chat models on common industry benchmarks. | 2024-07-23 | https://arxiv.org/abs/2407.21783 | https://github.com/meta-llama/llama3 | true | 15,000,000,000,000 | DecoderOnly | true | false | false | DPO | true | false | false | false | null | 0 | null | false | null | 0.579268 | 0.506579 | 0.54046 | 0.791946 | 0.771739 | 0.654762 | 0.755814 | 0.709091 | 0.7 | 0.419259 |
mixtral-8x7b-instruct | Mixtral-8x7b-Instruct | 47,000,000,000 | Mixtral is a sparse mixture-of-experts network. It is a decoder-only model where the feedforward block picks from a set of 8 distinct groups of parameters. At every layer, for every token, a router network chooses two of these groups (the “experts”) to process the token and combine their output additively. | 2023-12-11 | https://arxiv.org/abs/2401.04088 | https://huggingface.co/mistralai/Mixtral-8x7B-v0.1 | true | null | DecoderOnly | true | false | true | DPO | false | false | false | false | null | 0 | null | false | null | 0.42396 | 0.269737 | 0.535465 | 0.422819 | 0.554348 | 0.416667 | 0.47907 | 0.333333 | 0.325 | 0.26963 |
llama3.1-8b-instruct-T-one | Llama-3.1-8B-Instruct (Temperature 1.0) | 8,000,000,000 | The Meta Llama 3.1 collection of multilingual large language models (LLMs) is a collection of pretrained and instruction tuned generative models in 8B, 70B and 405B sizes (text in/text out). The Llama 3.1 instruction tuned text only models (8B, 70B, 405B) are optimized for multilingual dialogue use cases and outperform many of the available open source and closed chat models on common industry benchmarks. | 2024-07-23 | https://arxiv.org/abs/2407.21783 | https://github.com/meta-llama/llama3 | true | 15,000,000,000,000 | DecoderOnly | true | false | false | DPO | true | false | false | false | null | 1 | null | false | null | 0.461263 | 0.361842 | 0.523477 | 0.530201 | 0.478261 | 0.416667 | 0.576744 | 0.418182 | 0.4 | 0.32 |
gpt-4o | GPT-4o | null | GPT-4o is OpenAI's third major iteration of their popular large multimodal model, GPT-4, which expands on the capabilities of GPT-4 with Vision. | 2024-05-13 | null | null | true | null | DecoderOnly | false | false | null | null | true | false | false | false | null | 0 | null | false | null | 0.610832 | 0.559211 | 0.589411 | 0.805369 | 0.804348 | 0.75 | 0.755814 | 0.715152 | 0.75 | 0.441481 |
llama3-70b-instruct-T-one | Llama-3-70B-Instruct (Temperature 1.0) | 70,000,000,000 | Llama 3 models were trained on a text corpus of over 15T tokens. These models use a tokenizer with a vocabulary of 128K tokens. Additionally, improvements in the post-training procedures substantially reduced false refusal rates, improved alignment, and increased diversity in model responses. | 2024-04-18 | null | https://github.com/meta-llama/llama3 | true | 15,000,000,000,000 | DecoderOnly | true | false | false | DPO | true | false | false | false | null | 1 | null | false | null | 0.516499 | 0.375 | 0.53047 | 0.604027 | 0.684783 | 0.619048 | 0.632558 | 0.6 | 0.6 | 0.373333 |
paper-qa | PaperQA2 | null | PaperQA2 is a package for doing high-accuracy retrieval augmented generation (RAG) on PDFs or text files, with a focus on the scientific literature. We used PaperQA2 via the non-public API deployed by FutureHouse and the default settings (using Claude-3.5-Sonnet as summarizing and answer-generating LLM). | 2024-09-11 | https://storage.googleapis.com/fh-public/paperqa/Language_Agents_Science.pdf | https://github.com/Future-House/paper-qa | false | null | null | null | null | null | null | null | null | null | true | Paper Search, Gather Evidence, Generate Answer, Citation Traversal | 0 | null | false | null | 0.568867 | 0.460526 | 0.563437 | 0.724832 | 0.73913 | 0.690476 | 0.67907 | 0.678788 | 0.7 | 0.423704 |
gemma-1-1-7b-it | Gemma-1.1-7B-it | 7,000,000,000 | Gemma is a family of lightweight, open models built from the research and technology that Google used to create the Gemini models. Gemma models are well-suited for a variety of text generation tasks, including question answering, summarization, and reasoning. | 2024-02-21 | https://arxiv.org/abs/2403.08295 | https://github.com/google-deepmind/gemma | true | 6,000,000,000,000 | DecoderOnly | true | false | false | PPO | true | false | false | false | null | 0 | null | false | null | 0.192253 | 0.210526 | 0.004995 | 0.33557 | 0.413043 | 0.357143 | 0.37907 | 0.290909 | 0.375 | 0.22963 |
llama2-13b-chat | Llama-2-13B Chat | 13,000,000,000 | LLaMA, a collection of foundation language models ranging from 7B to 65B parameters. Llama was released by Meta AI. | 2023-07-18 | https://arxiv.org/abs/2302.13971 | https://huggingface.co/meta-llama/Llama-2-13b-chat-hf | false | 2,000,000,000,000 | DecoderOnly | true | false | false | null | true | false | false | false | null | 0 | null | false | null | 0.25538 | 0.092105 | 0.484515 | 0.114094 | 0.271739 | 0.095238 | 0.153488 | 0.151515 | 0.1 | 0.100741 |
llama3.1-8b-instruct | Llama-3.1-8B-Instruct | 8,000,000,000 | The Meta Llama 3.1 collection of multilingual large language models (LLMs) is a collection of pretrained and instruction tuned generative models in 8B, 70B and 405B sizes (text in/text out). The Llama 3.1 instruction tuned text only models (8B, 70B, 405B) are optimized for multilingual dialogue use cases and outperform many of the available open source and closed chat models on common industry benchmarks. | 2024-07-23 | https://arxiv.org/abs/2407.21783 | https://github.com/meta-llama/llama3 | true | 15,000,000,000,000 | DecoderOnly | true | false | false | DPO | true | false | false | false | null | 0 | null | false | null | 0.471664 | 0.394737 | 0.527473 | 0.503356 | 0.5 | 0.404762 | 0.581395 | 0.509091 | 0.45 | 0.325926 |
gemma-2-9b-it-T-one | Gemma-2-9B-it (Temperature 1.0) | 9,000,000,000 | Gemma is a family of lightweight, open models built from the research and technology that Google used to create the Gemini models. Gemma models are well-suited for a variety of text generation tasks, including question answering, summarization, and reasoning. | 2024-06-27 | https://arxiv.org/abs/2408.00118 | https://github.com/google-deepmind/gemma | true | 8,000,000,000,000 | DecoderOnly | true | false | false | PPO | true | false | false | false | null | 1 | null | false | null | 0.480273 | 0.289474 | 0.557443 | 0.557047 | 0.543478 | 0.5 | 0.546512 | 0.466667 | 0.475 | 0.342222 |
claude3.5-react | Claude-3.5 (Sonnet) React | null | Claude models are general purpose large language models. They use a transformer architecture and are trained via unsupervised learning, RLHF, and Constitutional AI (including both a supervised and Reinforcement Learning (RL) phase). Claude 3.5 was developed by Anthropic. | 2024-06-20 | null | null | true | null | null | false | false | null | null | true | false | false | true | ArXiV, Web search, Wikipedia, Wolfram alpha calculator, SMILES to IUPAC name and IUPAC name to SMILES converters | 0 | null | false | null | 0.624821 | 0.578947 | 0.599401 | 0.872483 | 0.804348 | 0.678571 | 0.837209 | 0.757576 | 0.8 | 0.408889 |
llama3.1-70b-instruct-T-one | Llama-3.1-70B-Instruct (Temperature 1.0) | 70,000,000,000 | The Meta Llama 3.1 collection of multilingual large language models (LLMs) is a collection of pretrained and instruction tuned generative models in 8B, 70B and 405B sizes (text in/text out). The Llama 3.1 instruction tuned text only models (8B, 70B, 405B) are optimized for multilingual dialogue use cases and outperform many of the available open source and closed chat models on common industry benchmarks. | 2024-07-23 | https://arxiv.org/abs/2407.21783 | https://github.com/meta-llama/llama3 | true | 15,000,000,000,000 | DecoderOnly | true | false | false | DPO | true | false | false | false | null | 1 | null | false | null | 0.511119 | 0.368421 | 0.535465 | 0.66443 | 0.695652 | 0.654762 | 0.553488 | 0.557576 | 0.55 | 0.38963 |
gemma-2-9b-it | Gemma-2-9B-it | 9,000,000,000 | Gemma is a family of lightweight, open models built from the research and technology that Google used to create the Gemini models. Gemma models are well-suited for a variety of text generation tasks, including question answering, summarization, and reasoning. | 2024-06-27 | https://arxiv.org/abs/2408.00118 | https://github.com/google-deepmind/gemma | true | 8,000,000,000,000 | DecoderOnly | true | false | false | PPO | true | false | false | false | null | 0 | null | false | null | 0.482425 | 0.315789 | 0.551449 | 0.543624 | 0.554348 | 0.52381 | 0.555814 | 0.484848 | 0.525 | 0.339259 |
llama2-70b-chat | Llama-2-70B Chat | 70,000,000,000 | LLaMA, a collection of foundation language models ranging from 7B to 65B parameters. Llama was released by Meta AI. | 2023-07-18 | https://arxiv.org/abs/2302.13971 | https://huggingface.co/meta-llama/Llama-2-70b-chat-hf | false | 2,000,000,000,000 | DecoderOnly | true | false | false | null | true | false | false | false | null | 0.01 | null | false | null | 0.266141 | 0.072368 | 0.487512 | 0.134228 | 0.217391 | 0.178571 | 0.146512 | 0.169697 | 0.125 | 0.136296 |
llama3-8b-instruct | Llama-3-8B-Instruct | 8,000,000,000 | Llama 3 models were trained on a text corpus of over 15T tokens. These models use a tokenizer with a vocabulary of 128K tokens. Additionally, improvements in the post-training procedures substantially reduced false refusal rates, improved alignment, and increased diversity in model responses. | 2024-04-18 | null | https://github.com/meta-llama/llama3 | true | 15,000,000,000,000 | DecoderOnly | true | false | false | DPO | true | false | false | false | null | 0 | null | false | null | 0.455524 | 0.407895 | 0.515485 | 0.442953 | 0.48913 | 0.416667 | 0.562791 | 0.369697 | 0.6 | 0.324444 |
gemma-1-1-7b-it-T-one | Gemma-1.1-7B-it (Temperature 1.0) | 7,000,000,000 | Gemma is a family of lightweight, open models built from the research and technology that Google used to create the Gemini models. Gemma models are well-suited for a variety of text generation tasks, including question answering, summarization, and reasoning. | 2024-02-21 | https://arxiv.org/abs/2403.08295 | https://github.com/google-deepmind/gemma | true | 6,000,000,000,000 | DecoderOnly | true | false | false | PPO | true | false | false | false | null | 1 | null | false | null | 0.190818 | 0.210526 | 0.008991 | 0.348993 | 0.413043 | 0.357143 | 0.372093 | 0.30303 | 0.375 | 0.216296 |
galactica_120b | Galactica-120b | 120,000,000,000 | Galactica is a large language model developed by Facebook. It is a transformer-based model trained on a large corpus of scientific data. | 2022-11-01 | https://galactica.org/paper.pdf | https://huggingface.co/facebook/galactica-120b | false | 450,000,000,000 | DecoderOnly | true | false | false | null | false | true | false | false | null | 0 | 4 | false | null | 0.015067 | 0 | 0 | 0.046053 | 0.053191 | 0 | 0.011338 | 0.055866 | 0 | 0.023188 |
llama3-8b-instruct-T-one | Llama-3-8B-Instruct (Temperature 1.0) | 8,000,000,000 | Llama 3 models were trained on a text corpus of over 15T tokens. These models use a tokenizer with a vocabulary of 128K tokens. Additionally, improvements in the post-training procedures substantially reduced false refusal rates, improved alignment, and increased diversity in model responses. | 2024-04-18 | null | https://github.com/meta-llama/llama3 | true | 15,000,000,000,000 | DecoderOnly | true | false | false | DPO | true | false | false | false | null | 1 | null | false | null | 0.4566 | 0.401316 | 0.52048 | 0.436242 | 0.543478 | 0.452381 | 0.551163 | 0.345455 | 0.625 | 0.324444 |
gemini-pro | Gemini-Pro | null | Gemini models are built from the ground up for multimodality: reasoning seamlessly across text, images, audio, video, and code. | 2024-06-07 | https://arxiv.org/abs/2403.05530 | null | true | null | DecoderOnly | false | false | null | null | true | false | false | false | null | 0 | null | false | null | 0.452654 | 0.388158 | 0.5005 | 0.483221 | 0.467391 | 0.5 | 0.567442 | 0.448485 | 0.475 | 0.308148 |
gpt-4 | GPT-4 | null | GPT-4 is a large multimodal model released by OpenAI to succeed GPT-3.5 Turbo. It features a context window of 32k tokens. | 2023-03-14 | https://arxiv.org/abs/2303.08774 | null | true | null | DecoderOnly | false | false | null | null | null | false | false | false | null | 0 | null | false | null | 0.412841 | 0.427632 | 0.163836 | 0.697987 | 0.695652 | 0.607143 | 0.67907 | 0.642424 | 0.7 | 0.41037 |
llama3-70b-instruct | Llama-3-70B-Instruct | 70,000,000,000 | Llama 3 models were trained on a text corpus of over 15T tokens. These models use a tokenizer with a vocabulary of 128K tokens. Additionally, improvements in the post-training procedures substantially reduced false refusal rates, improved alignment, and increased diversity in model responses. | 2024-04-18 | null | https://github.com/meta-llama/llama3 | true | 15,000,000,000,000 | DecoderOnly | true | false | false | DPO | true | false | false | false | null | 0 | null | false | null | 0.517934 | 0.414474 | 0.532468 | 0.604027 | 0.663043 | 0.630952 | 0.632558 | 0.593939 | 0.625 | 0.368889 |
phi-3-medium-4k-instruct | Phi-3-Medium-4k-Instruct | 14,000,000,000 | The Phi-3-Medium-4K-Instruct is a 14B parameters, lightweight, state-of-the-art open model trained with the Phi-3 datasets that includes both synthetic data and the filtered publicly available websites data with a focus on high-quality and reasoning dense properties. The model belongs to the Phi-3 family with the Medium version in two variants 4K and 128K which is the context length (in tokens) that it can support. | 2024-05-21 | https://arxiv.org/abs/2404.14219 | https://huggingface.co/microsoft/Phi-3-medium-4k-instruct | false | 4,800,000,000,000 | DecoderOnly | true | false | false | null | true | false | false | false | null | 0 | null | false | null | 0.474534 | 0.342105 | 0.532468 | 0.47651 | 0.630435 | 0.547619 | 0.560465 | 0.460606 | 0.55 | 0.331852 |
o1-preview | o1-preview | null | o1 is trained with reinforcement learning and chain-of-thought reasoning to improve safety, robustness, and reasoning capabilities. The family includes o1-preview and o1-mini versions. | 2024-09-12 | https://cdn.openai.com/o1-system-card-20240917.pdf | null | true | null | DecoderOnly | false | false | null | null | true | false | false | false | null | 1 | null | true | medium | 0.643472 | 0.625 | 0.563437 | 0.932886 | 0.902174 | 0.72619 | 0.830233 | 0.848485 | 0.85 | 0.475556 |
claude3 | Claude-3 (Opus) | null | Claude models are general purpose large language models. They use a transformer architecture and are trained via unsupervised learning, RLHF, and Constitutional AI (including both a supervised and Reinforcement Learning (RL) phase). Claude 3 was developed by Anthropic. | 2024-03-04 | https://www-cdn.anthropic.com/de8ba9b01c9ab7cbabf5c33b80b7bbc618857627/Model_Card_Claude_3.pdf | null | true | null | null | false | false | null | null | true | false | false | false | null | 0 | null | false | null | 0.569584 | 0.467105 | 0.565435 | 0.765101 | 0.793478 | 0.630952 | 0.695349 | 0.648485 | 0.7 | 0.41037 |
gpt-3.5-turbo | GPT-3.5 Turbo | null | GPT-3.5 Turbo, developed by OpenAI, features a context window of 4096 tokens. | 2023-11-06 | null | null | true | null | DecoderOnly | false | false | null | null | true | false | false | false | null | 0 | null | false | null | 0.466284 | 0.381579 | 0.534466 | 0.489933 | 0.543478 | 0.47619 | 0.588372 | 0.4 | 0.4 | 0.30963 |
claude2 | Claude-2 | null | Claude models are general purpose large language models. They use a transformer architecture and are trained via unsupervised learning, RLHF, and Constitutional AI (including both a supervised and Reinforcement Learning (RL) phase). Claude 2 was developed by Anthropic. | 2023-07-11 | https://www-cdn.anthropic.com/bd2a28d2535bfb0494cc8e2a3bf135d2e7523226/Model-Card-Claude-2.pdf | null | true | null | null | false | false | null | null | true | false | false | false | null | 0 | null | false | null | 0.473458 | 0.375 | 0.511489 | 0.503356 | 0.608696 | 0.464286 | 0.593023 | 0.50303 | 0.475 | 0.331852 |
README.md exists but content is empty.
- Downloads last month
- 536
Size of downloaded dataset files:
111 kB
Size of the auto-converted Parquet files:
111 kB
Number of rows:
170