WizardLM: An Instruction-following LLM Using Evol-Instruct

Empowering Large Pre-Trained Language Models to Follow Complex Instructions

News

• 🔥 We released 7B version of WizardLM. Checkout the paper and demo, demo_bak.
• 📣 We are looking for highly motivated students to join us as interns to create more intelligent AI together. Please contact [email protected]

Although on our complexity-balanced test set, WizardLM-7B outperforms ChatGPT in the high-complexity instructions, it still lag behind ChatGPT on the entire test set, and we also consider WizardLM to still be in a baby state. This repository will continue to improve WizardLM, train on larger scales, add more training data, and innovate more advanced large-model training methods.

Note for demo usage: Please use demo and demo_bak as evenly as possible to prevent one of them from being too heavy and responding slowly. We only recommend using English to experience our model. Support for other languages will be introduced in the future. The demo currently only supports single-turn conversation.

Overview of Evol-Instruct

Evol-Instruct is a novel method using LLMs instead of humans to automatically mass-produce open-domain instructions of various difficulty levels and skills range, to improve the performance of LLMs.

Contents

1. Online Demo

2. Training Data

3. WizardLM Weights

4. Fine-tuning

5. Inference

6. Evaluation

7. Citation

8. Disclaimer

Online Demo

We will provide our latest models for you to try for as long as possible. If you find a link is not working, please try another one. At the same time, please try as many real-world and challenging problems that you encounter in your work and life as possible. We will continue to evolve our models with your feedbacks.

Demo Link

Demo Backup

Training Data

alpaca_evol_instruct_70k.json contains 70K instruction-following data generated from Evol-Instruct. We used it for fine-tuning the WizardLM model. This JSON file is a list of dictionaries, each dictionary contains the following fields:

• instruction: str, describes the task the model should perform. Each of the 70K instructions is unique.
• output: str, the answer to the instruction as generated by gpt-3.5-turbo.

WizardLM Weights

We release [WizardLM] weights as delta weights to comply with the LLaMA model license. You can add our delta to the original LLaMA weights to obtain the WizardLM weights. Instructions:

1. Get the original LLaMA weights in the huggingface format by following the instructions here.
2. Please download our delta model at the following link
3. Use the following scripts to get WizardLM weights by applying our delta:

python src/weight_diff_wizard.py recover --path_raw <path_to_step_1_dir> --path_diff <path_to_step_2_dir> --path_tuned <path_to_store_recovered_weights>

Fine-tuning

We fine-tune WizardLM using code from Llama-X. We fine-tune LLaMA-7B with the following hyperparameters:

Hyperparameter	LLaMA-7B
Batch size	64
Learning rate	2e-5
Epochs	3
Max length	2048
Warmup step	2
LR scheduler	cosine

To reproduce our fine-tuning of WizardLM, please follow the following steps:

1. According to the instructions of Llama-X, install the environment, download the training code, and deploy.
2. Replace the train.py with the train_freeform.py in our repo(src/train_freeform.py)
3. Execute the following training command:

deepspeed train_freeform.py \
    --model_name_or_path /path/to/llama-7B/hf \
    --data_path /path/to/alpaca_evol_instruct_70k.json \
    --output_dir /path/to/wizardlm-7B/hf/ft \
    --num_train_epochs 3 \
    --model_max_length 2048 \
    --per_device_train_batch_size 8 \
    --per_device_eval_batch_size 1 \
    --gradient_accumulation_steps 1 \
    --evaluation_strategy "no" \
    --save_strategy "steps" \
    --save_steps 800 \
    --save_total_limit 3 \
    --learning_rate 2e-5 \
    --warmup_steps 2 \
    --logging_steps 2 \
    --lr_scheduler_type "cosine" \
    --report_to "tensorboard" \
    --gradient_checkpointing True \
    --deepspeed configs/deepspeed_config.json \
    --fp16 True

Inference

We provide the decoding script for WizardLM, which reads a input file and generates corresponding responses for each sample, and finally consolidates them into an output file.

You can specify base_model, input_data_path and output_data_path in src\inference_wizardlm.py to set the decoding model, path of input file and path of output file. The decoding command:

python src\inference_wizardlm.py

Evaluation

To evaluate Wizard, we conduct human evaluation on the inputs from our human instruct evaluation set WizardLM_testset.jsonl . This evaluation set was collected by the authors and covers a diverse list of user-oriented instructions including difficult Coding Generation & Debugging, Math, Reasoning, Complex Formats, Academic Writing, Extensive Disciplines, and so on. We performed a blind pairwise comparison between Wizard and baselines. Specifically, we recruit 10 well-educated annotators to rank the models from 1 to 5 on relevance, knowledgeable, reasoning, calculation and accuracy.

WizardLM achieved significantly better results than Alpaca and Vicuna-7b.

In the high-difficulty section of our test set (difficulty level >= 8), WizardLM even outperforms ChatGPT, with a win rate 7.9% larger than Chatgpt (42.9% vs. 35.0%). This indicates that our method can significantly improve the ability of large language models to handle complex instructions.

Citation

Please cite the repo if you use the data or code in this repo.

@misc{xu2023wizardlm,
      title={WizardLM: Empowering Large Language Models to Follow Complex Instructions}, 
      author={Can Xu and Qingfeng Sun and Kai Zheng and Xiubo Geng and Pu Zhao and Jiazhan Feng and Chongyang Tao and Daxin Jiang},
      year={2023},
      eprint={2304.12244},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

Disclaimer

The resources, including code, data, and model weights, associated with this project are restricted for academic research purposes only and cannot be used for commercial purposes. The content produced by any version of WizardLM 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.