THETA-Bench evaluation artifacts
Private collection point for evaluation artifacts produced on multiple servers. This repository is separate from the THETA training dataset and model weights. It starts with collection instructions only. The presence of this repository does not mean that evaluation data has already been uploaded or validated.
Read AGENT_REQUEST.md for the full Korean collection request. Each uploader needs existing write access to this private dataset. Never upload credentials or change repository visibility as part of data collection.
Collection scope
Collect evaluations of checkpoints trained on the canonical 3,003-segment THETA simulation set: 1,200 L1/L2 demonstrations and 1,803 extracted L0 prefixes. The eight models are ACT, DP, GR00T N1.7, Psi0, pi0.5, Cosmos 3, DreamZero 5B, and FastWAM. Each has 18 target conditions: three scenarios, three execution levels and two S/R modes, totaling 144 model-condition cells before additional checkpoint variants. Keep ACT K=16 and K=100 separate.
Collect eligible accessible results even when another agent or an earlier session ran the evaluation. Bind the actual checkpoint to its training manifest; a folder name is insufficient. Preserve original evaluator/host/run identities separately from the collector identity. Record intermediate checkpoints, diagnostics and real-data adaptation separately from the main simulation results. Keep missing conditions explicit and deduplicate repeated copies of the same original episode.
Upload layout
runs/<agent_id>/<run_id>/<snapshot_id>/
raw/
<original evaluation directory hierarchy>
run_metadata.json
episode_index.jsonl
coverage_3003.csv
files.sha256
MISSING.md
UPLOAD_COMPLETE.json
Use unique agent, run and snapshot IDs. Preserve worker and attempt boundaries inside the bundle. Keep all retained success/failure videos, metrics, summaries, optional trajectory JSON/NPZ pairs, and the configurations that produced them. Missing records remain missing. Do not infer contact events from terminal success.
Only the collection maintainer edits root documentation or a shared index. Each agent writes only to its own new snapshot prefix. Upload finalized files; leave running files out of the snapshot and report their status.
Upload the payload first, verify its file list, sizes and hashes at a pinned
Hub revision, and then upload UPLOAD_COMPLETE.json separately. The marker must
identify the data revision, manifest hash, file count, bytes, and run status.
Exclude the marker itself from the data manifest. Upload completion and
evaluation completion are separate states. Marked snapshots are immutable;
additional data goes into a new snapshot. Retries before completion use the same
unchanged files and prefix. Do not delete or overwrite other agents' files.
Using the collected data
Download a completed snapshot at its reported final commit and verify its manifest before analysis. Deduplicate overlapping snapshots by original run/worker/attempt/episode identity, preserving separate actual attempts.
The current THETA export_rollouts adapter imports outcomes and MP4s into
robot-traj-visualizer. Native measured trajectory JSON/NPZ must also be retained;
numeric body-pose plots require a converter or visualizer extension. Measured
joint positions are not executed policy actions. Training demonstrations and
model features are separate evidence types with their own provenance.
Storage usage counts toward the organization's Hub quota. Inspect planned bundle sizes and available storage before large transfers. This repository does not authorize storage-plan changes or public release of the collected artifacts.
- Downloads last month
- 89