CoolFace
Datasetpublic

novastar113/pusht_96_norm4_visual_nomarker_stopreq_aligned100k

PushT 96 Norm4 Visual Nomarker Stop-Required Aligned 100k This repository contains the exact plain PushT dataset used for the 2026-06-04 ordered SFT control run. Archive: pusht_96_norm4_visual_nomarker_stopreq_aligned100k_20260604_ordered.tar.gz Contents after extraction: data/train/: 100000 plain JSONL records in 8 gzip shards. data/test/: 100 plain JSONL records in 1 gzip shard. metadata/train_alignment_manifest.jsonl metadata/test_alignment_manifest.jsonl… See the full description on the dataset page: https://huggingface.co/datasets/novastar113/pusht_96_norm4_visual_nomarker_stopreq_aligned100k.

sourceHugging Faceotherupdated 4mo agoView on Hugging Face
0likes6downloads
Dataset Card

PushT 96 Norm4 Visual Nomarker Stop-Required Aligned 100k

This repository contains the exact plain PushT dataset used for the 2026-06-04 ordered SFT control run.

Archive:

  • pusht_96_norm4_visual_nomarker_stopreq_aligned100k_20260604_ordered.tar.gz

Contents after extraction:

  • data/train/: 100000 plain JSONL records in 8 gzip shards.
  • data/test/: 100 plain JSONL records in 1 gzip shard.
  • metadata/train_alignment_manifest.jsonl
  • metadata/test_alignment_manifest.jsonl
  • metadata/alignment_summary.json
  • metadata/visual_init_hash_audit_summary.json

Data contract:

  • Source: novastar112/pusht_96_norm4_visual_nomarker
  • Prompt rewrite: stop-required wording added to the plain PushT prompt.
  • CoT alignment: same selected 100000 train and 100 test source records, same order, as the paired CoT dataset.
  • Images: unchanged 96x96 JPEG base64 payloads.
  • Loader contract: ordered non-streaming JSONL path with __bagel_order_key.

Visual init-state hash audit:

  • Train rows: 100000, unique visual init hashes: 100000.
  • Test rows: 100, unique visual init hashes: 100.
  • Train internal duplicates: 0.
  • Test internal duplicates: 0.
  • Train/test visual init-state overlap: 0.

The visual init-state hash uses only agent_position, block_position, block_angle, and goal_pose, excluding episode seed, metadata, history, actions, velocities, and timestamps.