language: en

datasets:

• 37 popular Python code repositories
• See princeton-nlp/SWE-bench train split
• See the make_datasets documentation on SWE-bench's GitHub for details on formatting input.

SWE-Llama

SWE-Llama are variants of the CodeLlama model fine-tuned on software engineering tasks extracted from real-world GitHub issues and pull requests. They were introduced and evaluated on the SWE-bench benchmark in this paper.

Model Details

• Architecture: Transformer, based on CodeLlama architecture
• Parameters: 7 billion for SWE-Llama-7b, 13 billion for SWE-Llama-13b
• Objective: Generating patches to resolve GitHub issues, conditioned on issue description and code context

Training Data

SWE-Llama was fine-tuned on 19,000 issues and pull requests collected from 37 popular Python code repositories on GitHub, disjoint from those used in SWE-bench.

Training Procedure

• Fine-tuned only the attention matrices using LoRA method
• Trained for 4 epochs with a batch size of 32
• Selected best checkpoint based on validation perplexity

Evaluation Results

When evaluated on the SWE-bench benchmark:

• SWE-Llama-7b achieved 3.0% issue resolution rate using oracle context retrieval
• SWE-Llama-13b achieved 4.0% issue resolution rate using oracle context retrieval

BibTeX Entry

@misc{jimenez2023swebench,
      title={SWE-bench: Can Language Models Resolve Real-World GitHub Issues?}, 
      author={Carlos E. Jimenez and John Yang 
        and Alexander Wettig and Shunyu Yao 
        and Kexin Pei and Ofir Press and Karthik Narasimhan},
      year={2023},
      eprint={2310.06770},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}