Qwen2.5-Coder-32B-CL

Model Card for LocAgent

This model is described in the paper LocAgent: Graph-Guided LLM Agents for Code Localization. LocAgent uses a graph-based code representation to enable LLMs to perform accurate code localization, significantly improving accuracy compared to existing methods. Notably, the fine-tuned Qwen-2.5-Coder-Instruct-32B model achieves near state-of-the-art performance with a substantial cost reduction.

Code: https://github.com/gersteinlab/LocAgent

How to Use

LocAgent involves two main steps: graph indexing and code localization.

1. Graph Indexing (Optional but Recommended): For efficient batch processing, pre-generate graph indexes for your codebase using dependency_graph/batch_build_graph.py. This script parses the codebase into a graph representation. See the Github README for detailed command-line arguments and setup instructions. Example:

python dependency_graph/batch_build_graph.py \
       --dataset 'czlll/Loc-Bench' \
       --split 'test' \
       --num_processes 50 \
       --download_repo

2. Code Localization: Use auto_search_main.py to perform code localization. This script leverages LLMs to search and locate relevant code entities within the pre-generated graph indexes. See the Github README for detailed command-line arguments and environment variable setup. Example:

python auto_search_main.py \
   --dataset 'czlll/SWE-bench_Lite' \
   --split 'test' \
   --model 'azure/gpt-4o' \
   --localize \
   --merge \
   --output_folder $result_path/location \
   --eval_n_limit 300 \
   --num_processes 50 \
   --use_function_calling \
   --simple_desc

3. Evaluation: After localization, evaluate the results using evaluation.eval_metric.evaluate_results. An example Jupyter Notebook is provided in evaluation/run_evaluation.ipynb.

Citation

@article{chen2025locagent,
  title={LocAgent: Graph-Guided LLM Agents for Code Localization},
  author={Chen, Zhaoling and Tang, Xiangru and Deng, Gangda and Wu, Fang and Wu, Jialong and Jiang, Zhiwei and Prasanna, Viktor and Cohan, Arman and Wang, Xingyao},
  journal={arXiv preprint arXiv:2503.09089},
  year={2025}
}