Bug localization for real repositories
RepoTrace turns a bug report into a ranked list of source files. It combines repository-aware data engineering, lexical retrieval, and a fine-tuned CodeBERT experiment.
$ repotrace analyze apache/airflow --issue 36219
First changed file at rank 1. The first result matches the file changed by the fixing commit.
These examples come from repository-separated evaluation data. Choose a case to see what the system ranked and what the eventual fix changed.
$ repotrace analyze apache/airflow --issue 36219
First changed file at rank 1. The first result matches the file changed by the fixing commit.
Training and evaluation use the exact source snapshot that existed before each bug was fixed. Repositories never cross data splits.
Check out the repository before the fixing commit.
Filter and split source files into searchable chunks.
Use BM25 to select 50 distinct candidate files.
Score candidates with the fine-tuned CodeBERT experiment.
CodeBERT improved validation results, then failed to generalize to the held-out repositories. RepoTrace reports that result instead of hiding it.
The model experiment sits inside a tested application, not a notebook-only prototype.
Git history reconstruction, source filtering, hard-negative mining, and repository-disjoint splits.
CodeBERT fine-tuning, weight selection on validation, held-out ranking metrics, and error analysis.
FastAPI inference, Next.js interface, Docker packaging, request limits, and CI checks.