Compact-Bench

Evaluation inputs for Compact-Bench. The runner downloads the required files from this repository automatically.

Track	What it measures	Data
Task Preservation	Whether a compacted history supports completing a repository repair	35 tasks, each with a frozen history and repository snapshot
Recurring-Detail Retention	Literal retention of details needed again later	146 sessions (143 real, 3 simulated), with 9,330 annotated atoms
Production-Chain Retention	Detail retention across recorded compaction boundaries	106 real chains and 224 boundaries

Run

Clone the code, configure model credentials and endpoints using .env.example, then run:

python run.py --model YOUR_SUMMARIZER_MODEL

The full evaluation needs Linux, Python 3.9+, git, and patch; a Docker Compose entry point is provided in the code repository. To run the deterministic production-chain evaluation without model credentials:

python run.py --track production

No manual dataset download or datasets installation is required. Use --revision to select a specific HF commit or release tag.

Files

e1/instances.json   Task metadata, issue statements, patches and test lists
e1/H3/             35 frozen histories, one JSON file per task
e1/d3snap/         35 matching working-directory snapshots, in tar.gz format
e1/prepass24.txt    Historical cohort identifiers (see below)
e2/sessions.jsonl  Histories and recurring-detail annotations
e3/chains.jsonl    Recorded summaries and following-request annotations
release.json      Public release metadata
SHA256SUMS        File checksums

The filenames retain the original internal track IDs. They map to the three named tracks in the order shown above.

prepass24.txt is a historical selection of 24 instances, including the two excluded pytest instances (pytest-dev__pytest-10051 and pytest-dev__pytest-7236). The current runner computes the PRE-pass denominator from the current run and omits these two instances from aggregate task rates.

Session records contain sample_id, client, history H, and atoms with kind and value. Chain records contain sess, client, n_bounds, and per_k; each boundary includes the recorded summary S_k and atom groups atomsF. These files contain nested tool messages with different content types; use the benchmark's JSON readers. The tabular dataset viewer is disabled.

Provenance and release processing

Task inputs derive from SWE-bench and the public repositories listed in NOTICE.md; each task records its upstream repository and base commit. Session and chain records come from coding-agent harness runs, with three simulated sessions included in the retention track. The five production compaction prompts are distributed in the code repository.

The public export replaces detected credential strings consistently across histories, summaries and labels. It retains the task histories and repository snapshots unchanged. This processing is credential redaction, not a claim of exhaustive anonymization. The original inputs are kept separately by the maintainer; release processing is implemented in scripts/prepare_data.py.

License

Compact-Bench's original contributions are provided under Apache-2.0. Third-party source files and other upstream material retain their applicable licenses and notices; see NOTICE.md and the license files inside the snapshots.