News

• 📢 We have improved our model and devleoped a new framework Xolver, It achievevs a new State-of-the-art results—98.1% on GSM8K, 94.4% on AIME’24, 93.7% on AIME’25, 99.8% on Math-500, and 91.6% on LiveCodeBench.
• 🎉 Our new paper CodeSim got accepted in NAACL 2025 Findings.
• 🎉 Our new paper MapCoder got accepted in ACL 2024.
• All our codebase in both MapCoder and Redcoder are open-sourced with MIT and Modified MIT License.
• See you at ACL, 2024, Bangkok, Thailand.

REDCODER (Retrieval augmentED CODe gEneration and summaRization)

This is repository for the SCODE-R retriever in the Retrieval Augmented Code Generation and Summarization paper.

If you find this paper or this code useful, please cite this paper:

@inproceedings{parvez2021retrieval,
  title = {Retrieval Augmented Code Generation and Summarization},
  author = {Parvez, Md Rizwan and Ahmad, Wasi Uddin and Chakraborty, Saikat and Ray, Baishakhi and Chang, Kai-Wei},
  booktitle = {EMNLP-Findings},
  year = {2021}
}

Our model has two parts. You can use them seperately as well.

• SCODE-R: Summary and Code Retriever. Please see instructions in ./SCODE-R.
• SCODE-G: Summary and Code Generator. Please see instructions in ./SCODE-G.

All REDCODER data/models/outputs together:

• Exclude retrieval candidate embeddings(too large)
• Exclude tokenized input to SCODE-G (by sentencepiece, we provide code and docs in SCODE-G directory. Please use them instead.)
• Please go through issues specially this issue
• Sample SCODE-R output: code to text valid split top 30 k retrievals
• Finetuned SCODE-R checkpoints:
  • Code2Text Python: Link
  • Text2Code Python: Link
  • Code2Text Java: Link
  • Text2Code Java: Link
• All the retrieval database: (a) one combined summary retrieval corpus for code2text for both python and Java (b) Java and Python code retrieval corpus for text2Code tasks: LINK