Choiszt/gpt6-rollout

GPT-6 Astra: complete saved HumanCLAW rollout artifacts

This dataset preserves all saved artifacts of the original 1,218-episode GPT-6 Astra low-thinking run on HumanCLAW-Bench, including intermediate text prompts and visible model responses, verifier outputs, trajectories, metrics, and both video views.

It is an archive of an existing experiment, not a new evaluation run.

Contents

Artifact	Count
Episodes / episode TAR archives	1,218
Perception/planning/action JSON outputs	73,706
Verifier JSON outputs	49,365
Episode metrics	1,218
Replay manifests	1,218
Before/after trajectory NPZ files	2,436
Ego/exo MP4 videos	2,436
Original saved files, including the aggregate metrics	130,380

The original files occupy 20,483,508,640 bytes (20.48 GB decimal). TAR archives also have container overhead. Files are grouped by episode to avoid publishing more than 130,000 individual files. Source file contents are unchanged.

episodes/<episode>.tar
  <episode>/rollout_00/
    stepNNN_percept_mid_low.json
    stepNNN_verifier.json
    metrics.json
    replay_manifest.json
    trajectory_before.npz
    trajectory_after.npz
    ego.mp4
    exo.mp4
run/
  metrics_summary.json
  runner.log
  exit_status.txt
  launcher.pid
metadata/
  episode_index.csv
  files_manifest.jsonl
  all_metrics.jsonl
  inventory.json
  security_audit.json

metadata/episode_index.csv gives archive paths, archive SHA-256 values, sizes, and per-episode output counts. files_manifest.jsonl records the size and SHA-256 of every original file and its archive location. all_metrics.jsonl makes all episode metrics accessible without downloading videos or trajectories. The original launcher PID is historical metadata, not a currently running process.

All archive members were read back and checked against their source SHA-256 hashes before publication. The publisher separately verifies all uploaded objects against Hugging Face's stored size and LFS SHA-256 or Git blob hash.

Download

# Complete dataset
hf download Choiszt/gpt6-rollout --repo-type dataset --local-dir gpt6-rollout

# Small analysis metadata only
hf download Choiszt/gpt6-rollout --repo-type dataset \
  --include 'metadata/*' 'run/metrics_summary.json' 'README.md' \
  --local-dir gpt6-rollout

# One episode (replace EPISODE with a value in episode_index.csv)
hf download Choiszt/gpt6-rollout 'episodes/EPISODE.tar' \
  --repo-type dataset --local-dir gpt6-rollout
mkdir -p extracted
tar -xf gpt6-rollout/episodes/EPISODE.tar -C extracted

The original folder layout is retained inside each TAR. Videos are also individually accessible in the previous video dataset.

Provenance and interpretation

• Run ID: gpt6_fullval1218_low_8gpu_64par_20260907.
• Model deployment recorded for this experiment: gpt-6-astra, low-thinking setting.
• Harness: HumanCLAW, experiment-recorded commit 13a8b6d0bf9cc1e458aff0e7442b526288325acb, profile paper_fullval_v1.
• Original headline results: FindSR 75.45%, NavSR@20cm 57.14%, InteractSR 46.57% (278/597 sit-task episodes).
• These are the original, pre-object-escape-fix outputs. The 39 subsequently diagnosed episode replays and the 39 fresh model reruns under corrected physics are not substituted or mixed into this archive.
• Original aggregate dDtb is approximately 4.637829 m, affected by simulator object escapes. The later approximately 0.657 m value comes from a separate hybrid recomputation (39 same-action corrected replays plus 1,179 unchanged episodes); it is not the raw metric of this archive and not a full new 1,218-episode run.
• Collision is a per-episode collision-step fraction averaged over episodes, not video-frame percentage or collision severity. Original mean: approximately 41.47%. Different body parts may overlap.
• Runtime paths in replay manifests are preserved for provenance; scene assets, checkpoints, and environment dependencies must be supplied separately to replay the trajectories.

What “all saved outputs” means

Per-step JSON records contain the saved text prompt and the visible, structured final response of each logical perception/planning/action or verifier stage; an error may be present for a failed stage. Explicit reasoning fields in these responses are retained as model outputs. These are not complete raw HTTP request/response dumps or a separate archive of per-call input image payloads.

The original harness did not persist every retry response or any provider-hidden reasoning trace. It also did not save per-frame contact traces, per-step target-pixel/geodesic logs, or the older collector's episode_log.json format. Those unavailable records cannot be recovered simply by uploading this directory. The current archive does not claim to contain them.

All source JSON and included run-log text were screened for credential patterns before publication. No .env, API credential files, authentication caches, SSH configuration, source scene meshes/textures, or model checkpoints are included. No source outputs were redacted or otherwise rewritten in the published archives; the security-audit metadata reports the scan scope and limits.