File size: 7,402 Bytes
b06da0c 10143be b06da0c 10143be b06da0c 10143be b06da0c 4a8aed7 b06da0c 4a8aed7 b06da0c 297af45 b06da0c 297af45 b06da0c 5cf6c03 b06da0c 93407a2 b06da0c 4a8aed7 b06da0c 5cf6c03 b06da0c 4a8aed7 b06da0c 4a8aed7 9533253 4a8aed7 b06da0c 297af45 b06da0c cb1d680 b06da0c 29bb3ca b06da0c 29bb3ca b06da0c c2bf727 b06da0c 9533253 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 |
---
language:
- en
- ko
pipeline_tag: text-generation
tags:
- pytorch
- llama
- causal-lm
- 42dot_llm
license: cc-by-nc-4.0
---
# 42dot-plm-1.3b
**42dot-PLM** is a pre-trained language model (PLM) developed by [**42dot**](https://42dot.ai/) and is a part of **42dot LLM** (large language model). 42dot-PLM is pre-trained using Korean and English text corpus and can be used as a foundation language model for several Korean and English natural language tasks. This repository contains a 1.3B-parameter version of the model.
## Model Description
### Hyperparameters
42dot-PLM is built upon a Transformer decoder architecture similar to the [LLaMA 2](https://ai.meta.com/research/publications/llama-2-open-foundation-and-fine-tuned-chat-models/) and its hyperparameters are listed below.
| Params | Layers | Attention heads | Hidden size | FFN size |
| -- | -- | -- | -- | -- |
| 1.3B | 24 | 32 | 2,048 | 5,632 |
### Pre-training
Pre-training took 8 days using 256 * NVIDIA A100 GPUs. Related settings are listed below.
| Params | Global batch size\* | Initial learning rate | Train iter.\* | Max length\* | Weight decay |
| -- | -- | -- | -- | -- | -- |
| 1.3B | 4.0M | 4E-4 | 1.4T | 4,096 | 0.1 |
(\* unit: tokens)
### Pre-training datasets
We used a set of publicly available text corpus, including:
- Korean: including [Jikji project](http://jikji.duckdns.org/), [mC4-ko](https://huggingface.co/datasets/mc4), [LBox Open](https://github.com/lbox-kr/lbox-open), [KLUE](https://huggingface.co/datasets/klue), [Wikipedia (Korean)](https://ko.wikipedia.org/) and so on.
- English: including [The Pile](https://github.com/EleutherAI/the-pile), [RedPajama](https://github.com/togethercomputer/RedPajama-Data), [C4](https://huggingface.co/datasets/c4) and so on.
### Tokenizer
The tokenizer is based on the Byte-level BPE algorithm. We trained its vocabulary from scratch using a subset of the pre-training corpus. For constructing a subset, 10M and 10M documents are sampled from Korean and English corpus respectively. The resulting vocabulary sizes about 50K.
### Zero-shot evaluations
We evaluate 42dot-PLM on a variety of academic benchmarks both in Korean and English. All the results are obtained using [lm-eval-harness](https://github.com/EleutherAI/lm-evaluation-harness/tree/polyglot) and models released on the Hugging Face Hub.
#### Korean (KOBEST)
<figure align="center">
<img src="https://huggingface.co/42dot/42dot-plm-1.3b/resolve/main/asset/plm_benchmark_ko.png" width="90%" height="90%"/>
</figure>
|Tasks / Macro-F1|[KoGPT2](https://github.com/SKT-AI/KoGPT2) <br>1.2B|[Polyglot-Ko](https://github.com/EleutherAI/polyglot) <br>1.3B|[XGLM](https://huggingface.co/facebook/xglm-1.7B) <br>1.7B|[PolyLM](https://huggingface.co/DAMO-NLP-MT/polylm-1.7b) <br>1.7B|42dot-PLM <br>1.3B|
|--------------|-----------|----------------|---------|-----------|------------------------|
|boolq |0.337 |0.355 |**0.502** |0.334 |0.369 |
|copa |0.67 |**0.721** |0.616 |0.513 |0.704 |
|hellaswag |0.404 |0.401 |0.374 |0.321 |**0.431** |
|sentineg |0.606 |0.679 |0.46 |0.382 |**0.69** |
|**average** |0.504 |0.539 |0.488 |0.388 |**0.549** |
#### English
<figure align="center">
<img src="https://huggingface.co/42dot/42dot-plm-1.3b/resolve/main/asset/plm_benchmark_en.png" width="90%" height="90%"/>
</figure>
| Tasks / Metric | MPT <br>1B | OPT <br>1.3B | XGLM <br>1.7B | PolyLM <br>1.7B | 42dot-PLM <br>1.3B |
| ---------------------- | ------ | -------- | --------- | ----------- | ------------------------ |
| anli_r1/acc | 0.309 | **0.341** | 0.334 | 0.336 | 0.325 |
| anli_r2/acc | 0.334 | 0.339 | 0.331 | 0.314 | **0.34** |
| anli_r3/acc | 0.33 | 0.336 | 0.333 | **0.339** | 0.333 |
| arc_challenge/acc | 0.268 | 0.234 | 0.21 | 0.198 | **0.288** |
| arc_challenge/acc_norm | 0.291 | 0.295 | 0.243 | 0.256 | **0.317** |
| arc_easy/acc | 0.608 | 0.571 | 0.537 | 0.461 | **0.628** |
| arc_easy/acc_norm | 0.555 | 0.51 | 0.479 | 0.404 | **0.564** |
| boolq/acc | 0.517 | 0.578 | 0.585 | 0.617 | **0.624** |
| hellaswag/acc | 0.415 | 0.415 | 0.362 | 0.322 | **0.422** |
| hellaswag/acc_norm | 0.532 | 0.537 | 0.458 | 0.372 | **0.544** |
| openbookqa/acc | **0.238** | 0.234 | 0.17 | 0.166 | 0.222 |
| openbookqa/acc_norm | 0.334 | 0.334 | 0.298 | 0.334 | **0.34** |
| piqa/acc | 0.714 | 0.718 | 0.697 | 0.667 | **0.725** |
| piqa/acc_norm | 0.72 | 0.724 | 0.703 | 0.649 | **0.727** |
| record/f1 | 0.84 | **0.857** | 0.775 | 0.681 | 0.848 |
| record/em | 0.832 | **0.849** | 0.769 | 0.674 | 0.839 |
| rte/acc | 0.541 | 0.523 | **0.559** | 0.513 | 0.542 |
| truthfulqa_mc/mc1 | 0.224 | 0.237 | 0.215 | **0.251** | 0.236 |
| truthfulqa_mc/mc2 | 0.387 | 0.386 | 0.373 | **0.428** | 0.387 |
| wic/acc | 0.498 | **0.509** | 0.503 | 0.5 | 0.502 |
| winogrande/acc | 0.574 | **0.595** | 0.55 | 0.519 | 0.583 |
| **avearge** | 0.479 | 0.482 | 0.452 | 0.429 | **0.492** |
## Limitations and Ethical Considerations
42dot-PLM shares a number of well-known limitations of other large language models (LLMs). For example, it may generate false and misinformative content since 42dot-PLM is also subject to [hallucination](https://en.wikipedia.org/wiki/Hallucination_(artificial_intelligence)). In addition, 42dot-PLM may generate toxic, harmful, and biased content due to the use of web-available training data. We strongly suggest that 42dot-PLM users should be aware of those limitations and take necessary steps to mitigate those issues.
## Disclaimer
The contents generated by 42dot LLM series ("42dot LLM") do not necessarily reflect the views or opinions of 42dot Inc. ("42dot"). 42dot disclaims any and all liability to any part for any direct, indirect, implied, punitive, special, incidental, or other consequential damages arising from any use of the 42dot LLM and its generated contents.
## License
The 42dot-PLM is licensed under the Creative Commons Attribution-NonCommercial 4.0 (CC BY-NC 4.0).
## Citation
```
@misc{42dot2023llm,
title={42dot LLM: A Series of Large Language Model by 42dot},
author={Woo-Jong Ryu and Sang-Kil Park and Jinwoo Park and Seongmin Lee and Yongkeun Hwang},
year={2023},
url = {https://github.com/42dot/42dot_LLM},
version = {pre-release},
}
``` |