README.md

---
license: apache-2.0
datasets:
- databricks/databricks-dolly-15k
language:
- en
metrics:
- rouge
base_model:
- facebook/opt-2.7b
pipeline_tag: text-generation
---
# MiniLLM-OPT-2.7B

[paper](https://arxiv.org/abs/2306.08543) | [code](https://github.com/microsoft/LMOps/tree/main/minillm)

**MiniLLM-OPT-2.7B** is an OPT-2.7B model distilled from [OPT-13B](https://huggingface.co/MiniLLM/teacher-OPT-13B) on [databricks-dolly-15k](https://huggingface.co/datasets/aisquared/databricks-dolly-15k)

<p align='left'>
    <img src="https://cdn-uploads.huggingface.co/production/uploads/624ac662102fcdff87be51b9/7hBWGZzYMJihCRQ70XoiQ.png" width="1000">
</p>

**Note**: MiniLLM requires an [SFT model](https://huggingface.co/MiniLLM/init-opt-2.7B) for initilization to perform the PPO optimization.

## Evaluation

We ask GPT-4 to give scores for the generated responses of MiniLLM. The prompts are taken from [databricks-dolly-15k](https://huggingface.co/datasets/aisquared/databricks-dolly-15k) (test set), [self-instruct](https://github.com/tatsu-lab/stanford_alpaca/blob/main/alpaca_data.json), and [vicuna](https://github.com/lm-sys/vicuna-blog-eval)

<p align='left'>
    <img src="https://cdn-uploads.huggingface.co/production/uploads/624ac662102fcdff87be51b9/rDXnaDbKH5mBYAmqGC-_a.png" width="1000">
</p>

## Baseline Models
+ [SFT w/o KD](https://huggingface.co/MiniLLM/SFT-opt-2.7B)
+ [KD](https://huggingface.co/MiniLLM/KD-opt-2.7B)
+ [SeqKD](https://huggingface.co/MiniLLM/SeqKD-opt-2.7B)


## Citation
```
@inproceedings{minillm,
  title={MiniLLM: Knowledge Distillation of Large Language Models},
  author={Gu, Yuxian and Dong, Li and Wei, Furu and Huang, Minlie},
  booktitle={Proceedings of ICLR},
  year={2024}
}
```