--- license: mit --- # Note: To comply with the LLaMA model license, we release Lion weights as _delta weights_. # Lion: Adversarial Distillation of Closed-Source Large Language Model

Lion

[📄 Paper] | [⌨ Github] | [💻 Demo]

### Tuned on 70k instruction-following data, Lion (7B) can achieve 95% capability of ChatGPT!

## News - **[May 26, 2023]** We released the model weights. Check out the [7B](https://huggingface.co/YuxinJiang/Lion) model! - **[May 25, 2023]** We released an [online demo](https://84bc5e1fdfbb976d51.gradio.live/), try our model here! - **[May 23, 2023]** We released the code for training and inference. ## Contents 1. [Overview](#overview) 2. [Online Demo](#online-demo) 3. [Recovering Lion weights](#recovering-lion-weights) 4. [Inference](#inference) 5. [Training Process](#training-process) 6. [Evaluation](#evaluation) 7. [Citation](#citation) 8. [Disclaimer](#disclaimer) ## Overview

The high-level overview of our adversarial distillation framework, where we craft a compact Student LLM based on a superior closed-source LLM that serves three roles: the **Teacher**, the **Referee**, and the **Generator**. From left to right, there are three stages in an iteration: 1) an _imitation_ stage to align the student’s response with the teacher’s response; 2) a _discrimination_ stage to identify hard samples; 3) a _generation_ stage to produce new hard samples for escalating the challenges presented to the student model. ## Online Demo We will provide our latest models for you to try for as long as possible. You may ask some questions to Lion and we are happy to hear your feedback! [**Demo Link**](https://84bc5e1fdfbb976d51.gradio.live/) (the UI interface is shown below)

Since the training data are English instruction-following examples, You'd better ask questions in English. However, we found Lion can also understand instructions in other languages to some extent. See the following case:

## Recovering Lion weights We release Lion weights as delta weights to comply with the LLaMA model license. - [Lion-7B (delta weights)](https://huggingface.co/YuxinJiang/Lion) You can add our delta to the original LLaMA weights to obtain the Lion weights. Instructions: 1. Get the original LLaMA weights in the huggingface format by following the instructions [here](https://huggingface.co/docs/transformers/main/model_doc/llama) 2. Please download our delta model from [Hugging Face](https://huggingface.co/YuxinJiang/Lion) 3. Use the following scripts to get Lion weights by applying our delta: ```bash python src/weight_diff.py recover --path_raw huggyllama/llama-7b --path_diff YuxinJiang/Lion --path_tuned ``` ## Inference For inference and training of Lion, please first install the requirements: ```bash pip install -r requirements.txt ``` We provide the decoding script for Lion, which reads a input file and generates corresponding responses for each sample, and finally consolidates them into an output file. ```bash python src/lion_inference.py \ --model_dir \ --data_dir \ --output_dir \ --num_gpus 8 ``` ## Training Process Below shows one iteration of our adversarial distillation framework. ### 1. Imitation Stage #### 1.1 Acquire the teacher's response on the Train Pool ```bash python src/chatgpt_inference.py \ -q \ -o \ --api_key ``` #### 1.2 Instruction-tuning the student based on the teacher’s response on the Train Pool Fine-tuning was conducted on on a machine with 8 A100 80G GPUs. ```bash torchrun --nproc_per_node=8 --master_port= src/train.py \ --model_name_or_path \ --data_path \ --bf16 True \ --output_dir result \ --num_train_epochs 3 \ --model_max_length 1024 \ --per_device_train_batch_size 1 \ --per_device_eval_batch_size 1 \ --gradient_accumulation_steps 8 \ --evaluation_strategy "no" \ --save_strategy "steps" \ --save_steps 500 \ --save_total_limit 1 \ --learning_rate 2e-5 \ --weight_decay 0. \ --warmup_ratio 0.03 \ --lr_scheduler_type "cosine" \ --logging_steps 1 \ --fsdp "full_shard auto_wrap" \ --fsdp_transformer_layer_cls_to_wrap 'LlamaDecoderLayer' \ --tf32 True ``` ### 2. Discrimination Stage #### 2.1 Acquire the teacher's response on the Cache Pool ```bash python src/chatgpt_inference.py \ -q \ -o \ --api_key ``` #### 2.2 Acquire the student's response on the Cache Pool ```bash python src/lion_inference.py \ --model_dir \ --data_dir \ --output_dir \ --num_gpus 8 ``` #### 2.3 Ask the referee to output two scores according to the respose quality of the teacher and the student ```bash python src/chatgpt_referee.py \ -a \ -o \ --api_key ``` #### 2.4 Discriminate hard instructions and easy instructions ```bash python src/discrimination.py \ --review_path \ --chatgpt_inference_path \ --lion_inference_path path_to_lion_inference_for_the_Cache_Pool \ --hard_save_path \ --easy_save_path ``` ### 3. Generation Stage Fill the `openai.api_key = ""` in [src/utils.py](https://github.com/YJiangcm/Lion/raw/master/src/utils.py). #### 3.1 Generate new hard instructions ```bash python -m src/generate_hard_instruction generate_instruction_following_data \ --seed_tasks_path \ --output_dir \ --num_instructions_to_generate 3000 ``` #### 3.2 Generate new easy instructions ```bash python -m src/generate_easy_instruction generate_instruction_following_data \ --seed_tasks_path \ --output_dir \ --num_instructions_to_generate 3000 ``` ## Evaluation ### Automatic Evaluation with GPT-4 we leverage GPT-4 to automatically rate the response quality (with scores from 1 to 10) between two models on 80 unseen [Vicuna-Instructions](https://github.com/lm-sys/FastChat/blob/main/fastchat/eval/table/question.jsonl). ChatGPT has been chosen as the reference model to estimate the relative capability of diverse LLMs against it. The relative score is reported in percentage, computed as the ratio of the sum of scores. **Relative Overall Response Quality**:

**Relative Response Quality of Diverse Task Categories**:

### Human Evaluation with Alignment Criteria We employ the alignment criteria proposed by Askell et al. (2021), which define that an assistant is considered aligned if it is characterized by being helpful, honest, and harmless (HHH). We performed a human evaluation on 252 [UserOriented-Instructions](https://github.com/yizhongw/self-instruct/blob/main/human_eval/user_oriented_instructions.jsonl). To estimate the won rate, we compare the frequency of won, tie, and lost between each pair of models below.

## Citation Please cite our paper if you use the code in this repo. ``` @article{DBLP:journals/corr/abs-2305-12870, author = {Yuxin Jiang and Chunkit Chan and Mingyang Chen and Wei Wang}, title = {Lion: Adversarial Distillation of Closed-Source Large Language Model}, journal = {CoRR}, volume = {abs/2305.12870}, year = {2023}, url = {https://doi.org/10.48550/arXiv.2305.12870}, doi = {10.48550/arXiv.2305.12870}, eprinttype = {arXiv}, eprint = {2305.12870}, biburl = {https://dblp.org/rec/journals/corr/abs-2305-12870.bib}, bibsource = {dblp computer science bibliography, https://dblp.org} } ``` ## Disclaimer ⚠️ Lion is intended and licensed for **research use ONLY**. Commercial use is **strictly prohibited**. The content produced by any version of Lion is influenced by uncontrollable variables such as randomness, and therefore, the accuracy of the output cannot be guaranteed by this project. This project does not accept any legal liability for the content of the model output, nor does it assume responsibility for any losses incurred due to the use of associated resources and output results.