Model Card for umd-zhou-lab/recycled-alpaca-7b-v2.0

This model is trained by fine-tuning llama-2 with recycled Alpaca data V2.

Model Details

Model Description

• Developed by: UMD Tianyi Zhou Lab
• Model type: An auto-regressive language model based on the transformer architecture
• License: Llama 2 Community License Agreement
• Finetuned from model: meta-llama/Llama-2-7b

Model Sources

• GitHub: Reflection-Tuning
• Paper: Reflection-Tuning: Data Recycling Improves LLM Instruction-Tuning
• Data: Coming soon

Uses

The primary use of this model is research on large language models and chatbots. The primary intended users of the model are researchers and hobbyists in natural language processing, machine learning, and artificial intelligence.

Training

We use the prompt from FastChat:

A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER: Hi ASSISTANT: Hello.</s>USER: Who are you? ASSISTANT: I am ...</s>......

Hyperparameter	Global Batch Size	Learning rate	Epochs	Max length	Weight decay	Warmup Rate
Recycled Models (7B)	128	2e-5	3	2048	0	0.03

Performance

The following table provides a comparison between our recycled models (V2) and baseline models on the AlpacaEval Leaderboard and Huggingface Open LLM Leaderboard.

The V2 Recycled Alpaca Data and WizardLM data, and the corresponding paper will be released soon.

	AlpacaEval		Avg	ARC	HellaSwag	MMLU	TruthfulQA		Model
Alpaca 7B	26.46		50.21	42.65	76.91	41.73	39.55		/
Recycled Alpaca 7B V2.0	79.58		56.05	54.01	78.07	46.69	45.41		[hf-Link]
WizardLM 7B	67.64		54.18	51.60	77.70	42.70	44.70		/
Recycled WizardLM 7B V2.0	83.48		56.79	54.78	77.86	45.63	48.91		[hf-Link]

Citation

Please consider citing our paper if you think our codes, data, or models are useful. Thank you!

@misc{li2023reflectiontuning,
      title={Reflection-Tuning: Data Recycling Improves LLM Instruction-Tuning}, 
      author={Ming Li and Lichang Chen and Jiuhai Chen and Shwai He and Heng Huang and Jiuxiang Gu and Tianyi Zhou},
      year={2023},
      eprint={2310.11716},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}