SupingSun/ReproBoundary-Execution-Trajectories

ReproBoundary Execution Trajectories

This dataset contains the released execution logs for ReproBoundary, a benchmark of agent capability boundaries in complex paper-to-code reproduction. The benchmark defines 47 fixed paper-reproduction tasks and runs each task with nine model-agent settings, producing 423 trajectories.

The task definitions, evaluator code, protocol templates, and website are kept in the ReproBoundary GitHub repository. This dataset repository stores the larger trajectory files and the small metadata needed to interpret and verify them.

Contents

• trajectories/TRAJECTORY_INDEX.csv links every run to its task, model-agent setting, status, log path, and SHA-256 hash.
• trajectories/runs/ contains one compressed agent.log.gz file for each of the 423 planned runs.
• task_pool/ contains the 47-task inventory and machine-readable task metadata.
• docs/REDACTION_POLICY.md explains how sensitive text was removed.
• DATASET_MANIFEST.json records release counts and source information.
• DATASET_SHA256SUMS fixes the bytes of every released file.
• scripts/verify_dataset.py verifies the dataset after download.

Download

This dataset is public. Install the official Hugging Face client in an isolated Python environment; no account or access token is required for download.

python3 -m venv .hf-venv
source .hf-venv/bin/activate
python -m pip install "huggingface_hub>=1.0,<2"

hf download SupingSun/ReproBoundary-Execution-Trajectories \
  --repo-type dataset \
  --revision 57e17e11f795c6df9f7806ad2262fd32897ec2e6 \
  --local-dir ReproBoundary-Execution-Trajectories

cd ReproBoundary-Execution-Trajectories

The download is about 399 MB and contains 423 gzip logs. Allow at least 1 GB of free space for the dataset, client environment, and temporary files.

Recorded run states

The archive contains:

• 418 completed evidence bundles;
• four logs whose final state remains unresolved; and
• one snapshot saved while a run was active.

These states are preserved as recorded. A log marked complete is not automatically a successful scientific reproduction.

Redaction and privacy

The logs are cleaned copies of the protected execution records. Detected credentials, authorization headers, cookies, private paths, service tokens, secret values in URLs, and private-key blocks were removed or replaced before release.

The dataset does not include raw provider requests, raw security telemetry, credentials, private machine records, datasets, model weights, saved training states, predictions, or complete third-party repositories.

Automatic redaction cannot prove that every possible sensitive string was found. The verification script checks the configured high-confidence patterns. Please report any suspected sensitive content without redistributing it.

Verify the download

Run from the dataset root:

python3 scripts/verify_dataset.py

Expected output:

verification OK: tasks=47 settings=9 trajectories=423

The check validates every indexed path and hash, gzip readability, release checksums, and configured sensitive-text patterns. A full check took about two minutes on the machine used for the release audit.

Inspect one trajectory

Run this from the downloaded dataset root. It checks the first indexed log and prints no more than 20 lines.

python3 - <<'PY'
import csv
import gzip
import hashlib
import itertools
from pathlib import Path

root = Path(".")
with (root / "trajectories/TRAJECTORY_INDEX.csv").open(encoding="utf-8") as f:
    row = next(csv.DictReader(f))

log = root / row["trajectory_path"]
digest = hashlib.sha256()
with log.open("rb") as f:
    for block in iter(lambda: f.read(1024 * 1024), b""):
        digest.update(block)

print("task:", row["task_id"])
print("state:", row["state_status"])
print("path:", log)
print("sha256 matches index:", digest.hexdigest() == row["released_log_sha256"])
print("first 20 lines:")
with gzip.open(log, "rt", encoding="utf-8", errors="replace") as f:
    for line in itertools.islice(f, 20):
        print(line.rstrip())
PY

Intended use

The dataset supports research on how coding agents plan, execute, and fail at fixed paper-reproduction tasks. It can also be used to reconstruct run chronology and study tool use under the released task definitions.

The logs do not include saved predictions, checkpoints, model weights, or training states. They therefore support chronology, tool-use, failure, and release-integrity analysis, but are not sufficient by themselves to recompute most task metrics. Metric replay requires the additional files and public dependencies named by the corresponding evaluation contract.

Do not treat a submitted metric or the final line of a log as proof of reproduction success. Required stages, delivered artifacts, independent metric recomputation, and safety evidence must be checked separately.

Leaderboard results are not included in this dataset.

License status

The project license is still under review. Public access does not by itself grant permission to redistribute files that do not carry their own license. Each source paper, dataset, model, and third-party project retains its original terms.