metadata

language:
  - en
  - ko
pipeline_tag: text-generation
tags:
  - pytorch
  - llama
  - causal-lm
  - 42dot_llm
license: cc-by-nc-4.0

42dot_LLM-SFT-1.3B

42dot LLM-SFT is a large language model (LLM) developed by 42dot which is trained to follow natural language instructions. 42dot LLM-SFT is a part of 42dot LLM, and derived from 42dot LLM-PLM by supervised fine-tuning (SFT). This repository contains a 1.3B-parameter version.

Model Description

Hyperparameters

As same as 42dot LLM-PLM, the model is built upon a Transformer decoder architecture similar to the LLaMA 2 and its hyperparameters are listed below.

Params	Layers	Attention heads	Hidden size	FFN size	Max. length*
1.3B	24	32	2,048	5,632	4,096

(* unit: tokens)

Supervised Fine-tuning

Fine-tuning took about 112 GPU hours (in NVIDIA A100). For the training dataset, we manually constructed (question or insturuction) and response pairs, which can either be single- or multi-turn.

Evaluation

Inspired by recent attempts like Vicuna, we evaluate 42dot LLM-SFT with other proprietary/open-sourced chatbots using GPT-4 for assessing various aspects of responses. The evaluation dataset consists of 121 prompts over 10 categories. The sample of the evaluation dataset and prompt template can be downloaded from our GitHub repo.

Baselines:
- ChatGPT using GPT-3.5-turbo and GPT-4
- Bard
- KORani-v2-13B

Model	GPT-3.5	GPT-4	Bard	KORani	42dot LLM-SFT
Params	Unknown	Unknown	Unknown	13B	1.3B

**Comparison between proprietary chatbots and 42dot LLM-SFT**

Limitations and Ethical Considerations

42dot LLM-SFT shares a number of well-known limitations of other LLMs. For example, it may generate false and misinformative content since 42dot LLM-SFT is also subject to hallucination. In addition, 42dot LLM-SFT may generate toxic, harmful, and biased content due to the use of web-available training data in the pre-training phase. We strongly suggest that 42dot LLM-SFT 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 LLM-SFT 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={42dot Inc.},
      year={2023},
      url = {https://github.com/42dot/42dot_LLM},
      version = {1.0.0},
}

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	32.39
ARC (25-shot)	36.09
HellaSwag (10-shot)	58.96
MMLU (5-shot)	25.51
TruthfulQA (0-shot)	39.98
Winogrande (5-shot)	58.41
GSM8K (5-shot)	0.68
DROP (3-shot)	7.11