🔥 News
- [2024/04/10] 🔥🔥🔥 WaveCoder repo, models released at 🤗 HuggingFace!
- [2023/12/26] WaveCoder paper released.
💡 Introduction
WaveCoder 🌊 is a series of large language models (LLMs) for the coding domain, designed to solve relevant problems in the field of code through instruction-following learning. Its training dataset was generated from a subset of code-search-net data using a generator-discriminator framework based on LLMs that we proposed, covering four general code-related tasks: code generation, code summary, code translation, and code repair.
| Model | HumanEval | MBPP(500) | HumanEvalFix(Avg.) | HumanEvalExplain(Avg.) |
|---|---|---|---|---|
| GPT-4 | 85.4 | - | 47.8 | 52.1 |
| 🌊 WaveCoder-DS-6.7B | 65.8 | 63.0 | 49.5 | 40.8 |
| 🌊 WaveCoder-Pro-6.7B | 74.4 | 63.4 | 52.1 | 43.0 |
| 🌊 WaveCoder-Ultra-6.7B | 79.9 | 64.6 | 52.3 | 45.7 |
🪁 Evaluation
Please refer to WaveCoder's GitHub repo for inference, evaluation, and training code.
# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM
tokenizer = AutoTokenizer.from_pretrained("microsoft/wavecoder-ultra-6.7b")
model = AutoModelForCausalLM.from_pretrained("microsoft/wavecoder-ultra-6.7b")
📖 License
This code repository is licensed under the MIT License. The use of DeepSeek Coder models is subject to the its License.
☕️ Citation
If you find this repository helpful, please consider citing our paper:
@article{yu2023wavecoder,
title={Wavecoder: Widespread and versatile enhanced instruction tuning with refined data generation},
author={Yu, Zhaojian and Zhang, Xin and Shang, Ning and Huang, Yangyu and Xu, Can and Zhao, Yishujie and Hu, Wenxiang and Yin, Qiufeng},
journal={arXiv preprint arXiv:2312.14187},
year={2023}
}
Note
WaveCoder models are trained on the synthetic data generated by OpenAI models. Please pay attention to OpenAI's terms of use when using the models and the datasets.