Lion: Adversarial Distillation of Closed-Source Large Language Model
[
Tuned on 70k instruction-following data, Lion (7B) can achieve 95% capability of ChatGPT!
News
We are currently working on training larger-sized versions (13B, 33B, and 65B if feasible). Thank you for your patience.
- [June 10, 2023] We released insturctions for addressing OOM during fine-tuning, check it in Training Process.
- [May 26, 2023] We released the model weights. Check out the 7B model!
- [May 25, 2023] We released an online demo, try our model here!
- [May 23, 2023] We released the code for training and inference.
Contents
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:
- an imitation stage to align the student’s response with the teacher’s response;
- a discrimination stage to identify hard samples;
- 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 (it will expire in 72 hours, so we regularly update the link)
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.
You can add our delta to the original LLaMA weights to obtain the Lion weights. Instructions:
- Get the original LLaMA weights in the huggingface format by following the instructions here
- Please download our delta model from Hugging Face
- Use the following scripts to get Lion weights by applying our delta:
python src/weight_diff.py recover --path_raw huggyllama/llama-7b --path_diff YuxinJiang/Lion --path_tuned <path_to_store_recovered_weights>Inference
For inference and training of Lion, please first install the requirements:
pip install -r requirements.txtWe 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. It can be run on a single machine with 16GB GPU.
python src/lion_inference.py \
--model_dir <path_to_hf_converted_lion_ckpt_and_tokenizer> \
--data_dir <path_to_input_json_file> \
--output_dir <path_to_output_json_file> \
--num_gpus 1Training Process
Below shows one iteration of our adversarial distillation framework.
1. Imitation Stage
1.1 Acquire the teacher's response on the Train Pool
python src/chatgpt_inference.py \
-q <path_to_json_file_for_the_Train_Pool> \
-o <path_to_chatgpt_inference_for_the_Train_Pool> \
--api_key <your_openai_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.
torchrun --nproc_per_node=8 --master_port=<your_random_port> src/train.py \
--model_name_or_path <path_to_hf_converted_ckpt_and_tokenizer> \
--data_path <path_to_chatgpt_inference_for_the_Train_Pool> \
--bf16 True \
--output_dir result \
--num_train_epochs 3 \
--model_max_length 1024 \
--per_device_train_batch_size 2 \
--per_device_eval_batch_size 2 \
--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 TrueAddressing OOM
Naively, fine-tuning a 7B model requires about 7 x 8 x 2 = 112 GB of VRAM. Commands given above enable parameter sharding, so no redundant model copy is stored on any GPU. If you'd like to further reduce the memory footprint, here are some options:
-
Turn on CPU offload for FSDP with
--fsdp "full_shard auto_wrap offload". This saves VRAM at the cost of longer runtime. -
In our experience, DeepSpeed stage-3 (with offload) can at times be more memory efficient than FSDP with offload. Here's an example to use DeepSpeed stage-3 with 8 GPUs with both parameter and optimizer offload:
deepspeed src/train_deepspeed.py \ --model_name_or_path <path_to_hf_converted_ckpt_and_tokenizer> \ --data_path <path_to_chatgpt_inference_for_the_Train_Pool> \ --output_dir result \ --num_train_epochs 3 \ --model_max_length 1024 \ --per_device_train_batch_size 16 \ --per_device_eval_batch_size 1 \ --gradient_accumulation_steps 1 \ --evaluation_strategy "no" \ --save_strategy "steps" \ --save_steps 600 \ --save_total_limit 1 \ --learning_rate 2e-5 \ --warmup_ratio 0.03 \ --logging_steps 1 \ --lr_scheduler_type "cosine" \ --report_to "tensorboard" \ --gradient_checkpointing True \ --deepspeed srcs/configs/deepspeed_config.json \ --fp16 True- The DeepSpeed library also provides some helpful functions to estimate memory usage.
-
LoRA fine-tunes low-rank slices of the query, key, and value embedding heads. This can reduce the total memory footprint from 112GB to about 7x4=28GB. We may release our re-implemention of this in the future, but for now the peft codebase can be a useful resource.
2. Discrimination Stage
2.1 Acquire the teacher's response on the Cache Pool
python src/chatgpt_inference.py \
-q <path_to_json_file_for_the_Cache_Pool> \
-o <path_to_chatgpt_inference_for_the_Cache_Pool> \
--api_key <your_openai_api_key>2.2 Acquire the student's response on the Cache Pool
python src/lion_inference.py \
--model_dir <path_to_hf_converted_lion_ckpt_and_tokenizer> \
--data_dir <path_to_json_file_for_the_Cache_Pool> \
--output_dir <path_to_lion_inference_for_the_Cache_Pool> \
--num_gpus 82.3 Ask the referee to output two scores according to the respose quality of the teacher and the student
python src/chatgpt_referee.py \
-a <path_to_chatgpt_inference_for_the_Cache_Pool> <path_to_lion_inference_for_the_Cache_Pool> \
-o <path_to_output_review_file> \
--api_key <your_openai_api_key>2.4 Discriminate hard instructions and easy instructions
python src/discrimination.py \
--review_path <path_to_output_review_file> \
--chatgpt_inference_path <path_to_chatgpt_inference_for_the_Cache_Pool> \
--lion_inference_path <path_to_lion_inference_for_the_Cache_Pool> \
--hard_save_path <path_to_identified_hard_instructions> \
--easy_save_path <path_to_identified_easy_instructions>3. Generation Stage
3.1 Generate new hard instructions
python -m src/generate_hard_instruction generate_instruction_following_data \
--seed_tasks_path <path_to_identified_hard_instructions> \
--output_dir <path_to_generated_hard_instructions> \
--num_instructions_to_generate 3000 \
--api_key <your_openai_api_key>3.2 Generate new easy instructions
python -m src/generate_easy_instruction generate_instruction_following_data \
--seed_tasks_path <path_to_identified_easy_instructions> \
--output_dir <path_to_generated_easy_instructions> \
--num_instructions_to_generate 3000 \
--api_key <your_openai_api_key>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. 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. 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},
timestamp = {Fri, 26 May 2023 11:29:33 +0200},
biburl = {https://dblp.org/rec/journals/corr/abs-2305-12870.bib},
bibsource = {dblp computer science bibliography, https://dblp.org}
}