Jack-Jieke-Wu/Paper-Writing-Exam-Trials

Paper-Writing Exam Agent Trials

This is a public archive of complete agent trials run on the public Jack-Jieke-Wu/Paper-Writing-Exam benchmark. It contains sanitized trajectories, agent outputs, final papers, Harbor results, verifier metrics, and agent-specific diagnostics.

The repository is public by design. Access to this repository does not permit redistribution of the benchmark, model outputs, or third-party source material. Every upload must still pass the local fail-closed exporter checks.

Layout

README.md
data/
├── trials.jsonl
├── events.jsonl
├── trials.schema.json
└── events.schema.json
artifacts/
└── <trial-id>.tar.gz
manifests/
├── release.json
└── <trial-id>.json

data/trials.jsonl has one record per Harbor trial. data/events.jsonl has one record per step derived from Harbor's native ATIF files (agent/trajectory*.json). The compressed archive contains the original ATIF trajectories, allowlisted submission, agent logs/checkpoints, Harbor result.json, and verifier evaluation.json. For Harbor v0.20.0, accepted native trajectory versions are ATIF-v1.0 through ATIF-v1.7; local file-based subagent references must resolve to included trajectory files.

Archives retain Harbor's native agent/, verifier/, steps/, and artifacts/ directories so single-step and multi-step trials remain reconstructable.

Saving A New Trial

harbor run writes a trial locally; it does not upload the result here. From the paperbench-harbor repository, export one completed Harbor trial with the sanitizing exporter:

python3 scripts/export_trial.py \
  --trial-dir /path/to/harbor-jobs/<job-name>/<trial-id> \
  --output-dir /path/to/Paper-Writing-Exam-Trials \
  --private-manifest /path/to/task/tests/private/source_manifest.json \
  --task-id pwb-0001 \
  --benchmark PaperWrite-Bench \
  --protocol short \
  --benchmark-hf-revision <immutable-benchmark-commit> \
  --harbor-repo-commit <paperbench-harbor-commit> \
  --agent-name codex \
  --agent-version <agent-version> \
  --integration-commit <integration-commit> \
  --model <provider>/<model> \
  --provider <provider> \
  --agent-config-file /path/to/non-secret-agent-config.json

The exporter creates artifacts/<trial-id>.tar.gz, manifests/<trial-id>.json, and appends a record to data/trials.jsonl. When the trial contains ATIF events, it also appends their one-event-per-step index to data/events.jsonl. It performs the credential/private-file checks before committing those outputs. Inspect the result, then upload the staging directory separately:

hf upload Jack-Jieke-Wu/Paper-Writing-Exam-Trials \
  /path/to/Paper-Writing-Exam-Trials . \
  --repo-type dataset \
  --exclude '.git/**' \
  --commit-message "Add trial <trial-id>"

Record the immutable commit SHA returned by this upload as TRIAL_DATASET_REVISION; use that value in the viewing commands below.

The full workflow, including how to choose --jobs-dir, is documented in the source repository's docs/trial-dataset.md.

View One Trial

Use the trial ID from data/trials.jsonl, then download its summary and manifest without downloading every archive:

TRIAL_ID=<trial-id>
TRIAL_DATASET_REVISION=<40-character-Hugging-Face-commit>
VIEW_DIR=/tmp/paper-writing-trial
mkdir -p "$VIEW_DIR"

hf download Jack-Jieke-Wu/Paper-Writing-Exam-Trials \
  data/trials.jsonl \
  "manifests/$TRIAL_ID.json" \
  --repo-type dataset \
  --revision "$TRIAL_DATASET_REVISION" \
  --local-dir "$VIEW_DIR"

If data/events.jsonl exists in the selected release, download it separately to inspect the one-event-per-step index.

To view the complete trajectory and final output, download and unpack the single archive:

hf download Jack-Jieke-Wu/Paper-Writing-Exam-Trials \
  "artifacts/$TRIAL_ID.tar.gz" \
  --repo-type dataset \
  --revision "$TRIAL_DATASET_REVISION" \
  --local-dir "$VIEW_DIR"
mkdir -p "$VIEW_DIR/unpacked"
tar -xzf "$VIEW_DIR/artifacts/$TRIAL_ID.tar.gz" -C "$VIEW_DIR/unpacked"

Then inspect:

$VIEW_DIR/unpacked/agent/trajectory*.json                 # single-step trajectory
$VIEW_DIR/unpacked/steps/*/agent/trajectory*.json          # multi-step trajectories
$VIEW_DIR/unpacked/artifacts/workspace/submission/         # single-step output
$VIEW_DIR/unpacked/steps/*/artifacts/workspace/submission/ # multi-step output
$VIEW_DIR/unpacked/harbor/result.json
$VIEW_DIR/unpacked/verifier/evaluation.json

data/events.jsonl is a one-event-per-step index derived from the original Harbor ATIF trajectory. The original trajectory and agent logs are preserved in the archive.

Provenance

Each trial references the benchmark without duplicating its task tree using:

• benchmark_hf_repo and immutable benchmark_hf_revision;
• task_id and task_checksum;
• harbor_repo_commit;
• agent_name, agent_version, and integration_commit;
• model/provider and a hash of non-secret configuration;
• Harbor reward, official metrics, timing, and artifact SHA-256.

Exports require the task's verifier-only tests/private/source_manifest.json and a SHA-256 hash of the non-secret agent configuration (or a configuration file from which that hash can be computed). The exporter compares source-material hashes before writing any output, and does not copy that manifest into this repository.

Security and privacy

The exporter rejects API keys, bearer tokens, cookies, credential files, host credentials, encoded credentials, solution/, tests/private/, eval_points.json, ground-truth papers, and unrelated host files. It scans a single temporary source snapshot, then derives the archive, event index, and manifest from that same snapshot. It does not upload task containers or publish from inside an evaluated container.

Do not upload raw environment files, authentication state, prompts containing secrets, or data that the benchmark license does not permit you to retain. Review every generated manifest before uploading it to this public repository.

Schemas are stored in data/*.schema.json. The local exporter and its tests are maintained at:

https://github.com/a-green-hand-jack/paperbench-harbor