pretty_name: TrajectoryTrace replication inputs
language:
- en
TrajectoryTrace: exact replication inputs
This dataset packages the local inputs used by the TrajectoryTrace project. It is separate from agenttraj-swebench-data: the local bug-triggering reference coverage differs from that older hosted coverage. The older dataset should not be substituted for these inputs when reproducing the current analyses.
Download and extract
Download download_dataset.py and requirements.txt from this dataset, or use code/download_project_data.py in the GitHub repository. Run from the project root:
python -m pip install -r requirements.txt
python download_dataset.py actions --list
python download_dataset.py actions
python download_dataset.py coverage
python download_dataset.py replay
python download_dataset.py trajectories
Use graphs for the additional graph data, or all for every component. Use --list to inspect compressed and extracted sizes first. Fully extracted inputs require about 219 GiB, plus space for compressed downloads. Cached archives are saved under data/.packages/.
The helper verifies both archive and extracted-file SHA-256 checksums, and restores the original data/ layout expected by the code. Existing differing files are preserved unless --force is specified. Use --revision COMMIT to pin downloads to a dataset commit.
Contents
archives/actions_all.tar.zst: parsed action log.archives/coverage.tar.zst: exact local reference coverage, buggy-version coverage, and associated metadata.archives/replay--*.tar.zst: replay coverage and original replay archives/metadata.archives/trajectory--*.tar.zst: original trajectories and evaluation records, grouped by configuration.archives/graph.tar.zst: graph inputs.inventory.json: package paths, sizes, counts, and archive checksums.archives/*.files.jsonl: original paths, byte sizes, and per-file SHA-256 checksums.SHA256SUMS.txt: archive checksums.
All input files are packaged from the local project without changing their contents. Local extraction markers and download-cache metadata are excluded. Raw trajectories can contain environment paths and recorded command outputs. The dataset preserves research evidence; labels and instrumentation have the limitations described in the manuscript. The manuscript is a working draft.
Project code, LaTeX, and compiled PDFs are on GitHub. Computed outputs are in the replication release. This dataset packages inputs, not claims of validated reproduction or causal effects. Third-party source and benchmark terms continue to apply; no new license is asserted here.
Verified package
The package contains 242,911 files in 63 archives, representing 218.82 GiB of original inputs and 2.09 GiB of compressed downloads. Every archived file was read back and checked against its SHA-256 manifest. Every hosted archive checksum was checked against the verified local package. See validation.json for the validation summary.
The download helper requires Python 3.11 or later.