Skip to content

High Accuracy and efficiency multi-task fine-tuning framework for Code LLMs

License

Notifications You must be signed in to change notification settings

paipai-Studio/MFTCoder

 
 

Repository files navigation

MFTCoder: Boosting Code LLMs with Multitask Fine-Tuning

stars forks License: MIT Open Issues

[中文] [English]

Contents

News

🔥🔥 [2023/11/07] MFTCoder Paper has been released on Arxiv, which discloses technique details of multi-task-fine-tuning.

🔥🔥 [2023/10/20] CodeFuse-QWen-14B has been released, achieving a pass@1 (greedy decoding) score of 48.8% on HumanEval, which gains 16% absolute improvement over the base model Qwen-14b

🔥🔥 [2023/09/27] CodeFuse-StarCoder-15B has been released, achieving a pass@1 (greedy decoding) score of 54.9% on HumanEval.

🔥🔥🔥 [2023/09/26]We are pleased to announce the release of the 4-bit quantized version of CodeFuse-CodeLlama-34B. Despite the quantization process, the model still achieves a remarkable 73.8% accuracy (greedy decoding) on the HumanEval pass@1 metric.

🔥🔥🔥 [2023/09/07]We released CodeFuse-CodeLlama-34B, which achieves the 74.4% Python Pass@1 (greedy decoding) and surpasses GPT4 (2023/03/15) and ChatGPT-3.5 on the HumanEval Benchmarks.

🔥🔥 [2023/08/26]We released MFTCoder which supports finetuning Code Llama, Llama, Llama2, StarCoder, ChatGLM2, CodeGeeX2, Qwen, and GPT-NeoX models with LoRA/QLoRA.

HumanEval Performance

Model HumanEval(Pass@1) Date
CodeFuse-CodeLlama-34B 74.4% 2023/09
CodeFuse-CodeLlama-34B-4bits 73.8% 2023/09
WizardCoder-Python-34B-V1.0 73.2% 2023/08
GPT-4(zero-shot) 67.0% 2023/03
PanGu-Coder2 15B 61.6% 2023/08
CodeFuse-StarCoder-15B 54.9% 2023/08
CodeLlama-34b-Python 53.7% 2023/08
CodeFuse-QWen-14B 48.8% 2023/10
CodeLlama-34b 48.8% 2023/08
GPT-3.5(zero-shot) 48.1% 2022/11
OctoCoder 46.2% 2023/08
StarCoder-15B 33.6% 2023/05
QWen-14B 32.3% 2023/10

Articles

MFT Arxiv paper

Introduction

High Accuracy and efficiency multi-task fine-tuning framework for Code LLMs.

CodeFuse-MFTCoder is an open-source project of CodeFuse for multitasking Code-LLMs(large language model for code tasks), which includes models, datasets, training codebases and inference guides. In MFTCoder, we released two codebases for finetuning Large Language Models:

  • mft_peft_hf is based on the HuggingFace Accelerate and deepspeed framework.
  • mft_atorch is based on the ATorch frameworks, which is a fast distributed training framework of LLM.

The aim of this project is to foster collaboration and share advancements in large language models, particularly within the domain of code development.

Frameworks

img.png

Highlights

Multi-task: Train models on multiple tasks while maintaining a balance between them. The models can even generalize to new, previously unseen tasks.

Multi-model: It integrates state-of-the-art open-source models such as gpt-neox, llama, llama-2, baichuan, Qwen, chatglm2, and more. (These finetuned models will be released in the near future.)

Multi-framework: It provides support for both HuggingFace Accelerate (with deepspeed) and ATorch.

Efficient fine-tuning: It supports LoRA and QLoRA, enabling fine-tuning of large models with minimal resources. The training speed meets the demands of almost all fine-tuning scenarios.

The main components of this project include:

  • Support for both SFT (Supervised FineTuning) and MFT (Multi-task FineTuning). The current MFTCoder achieves data balance among multiple tasks, and future releases will achieve a balance between task difficulty and convergence speed during training.
  • Support for QLoRA instruction fine-tuning, as well as LoRA fine-tuning.
  • Support for most mainstream open-source large models, particularly those relevant to Code-LLMs, such as Code-LLaMA, Starcoder, Codegeex2, Qwen, GPT-Neox, and more.
  • Support for weight merging between the LoRA adaptor and base models, simplifying the inference process.
  • Release of 2 high-quality code-related instruction fine-tuning datasets: Evol-instruction-66k and CodeExercise-Python-27k.
  • Release of 2 models: CodeFuse-13B and CodeFuse-CodeLlama-34B.

Requirements

To begin, ensure that you have successfully installed CUDA (version >= 11.4, preferably 11.7) along with the necessary drivers. Additionally, make sure you have installed torch (version 2.0.1).

Next, we have provided an init_env.sh script to simplify the installation of required packages. Execute the following command to run the script:

sh init_env.sh

If you require flash attention, please refer to the following link for installation instructions: https://github.com/Dao-AILab/flash-attention

Training

🚀 Huggingface accelerate + deepspeed Codebase for MFT(Multi-task Finetuning)

🚀 Atorch Codebase for MFT(Multi-task Finetuning)

Models

We are excited to release the following two CodeLLMs trained by MFTCoder, now available on Hugging Face:

Model Base Model Num of examples trained Batch Size Seq Length
🔥🔥🔥 CodeFuse-CodeLlama-34B CodeLlama-34b-Python 600k 80 4096
🔥🔥🔥 CodeFuse-CodeLlama-34B-4bits CodeLlama-34b-Python 4096
🔥🔥🔥 CodeFuse-StarCoder-15B Starcoder 600k 256 4096
🔥🔥🔥 CodeFuse-QWen-14B Qwen-14b 1100k 256 4096
🔥 CodeFuse-13B CodeFuse-13B 66k 64 4096

Datasets

We are also pleased to release two code-related instruction datasets, meticulously selected from a range of datasets to facilitate multitask training. Moving forward, we are committed to releasing additional instruction datasets covering various code-related tasks.

Dataset Description
⭐ Evol-instruction-66k Based on open-evol-instruction-80k, filter out low-quality, repeated, and similar instructions to HumanEval, thus get high-quality code instruction dataset.
⭐ CodeExercise-Python-27k python code exercise instruction dataset generated by chatgpt

Contributing

Contributions are welcome! If you have any suggestions, ideas, bug reports, or new model/feature supported, please open an issue or submit a pull request.

Citation

If you find our work useful or helpful for your R&D works, please feel free to cite our paper as below.

@article{mftcoder2023,
      title={MFTCoder: Boosting Code LLMs with Multitask Fine-Tuning}, 
      author={Bingchang Liu and Chaoyu Chen and Cong Liao and Zi Gong and Huan Wang and Zhichao Lei and Ming Liang and Dajun Chen and Min Shen and Hailian Zhou and Hang Yu and Jianguo Li},
      year={2023},
      journal={arXiv preprint arXiv},
      archivePrefix={arXiv},
      eprint={2311.02303}
}

Star-History

Star History Chart

About

High Accuracy and efficiency multi-task fine-tuning framework for Code LLMs

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 93.0%
  • C++ 6.4%
  • Other 0.6%