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)	HumanEval Fix(Avg.)	HumanEval Explain(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.

python

# 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.