tsilva/gradlab-breakout-c6d579da

Breakout checkpoint trajectories

1000 complete episodes, 9573810 transitions, 8767749 unique lossless WebP RGB images.

Trajectory schema version 1; machine-readable contract. The publication receipt and every Parquet shard declare the version and contract fingerprint.

Table schemas and brick annotations match the existing trajectories dataset. The full 210×160 RGB image includes the HUD. The original all view retains null episode split labels. Balanced train, validation, and test views are also available; their episode tables contain split assignments. Frame IDs are identifiers, never row offsets. source_frame_id and successor_frame_id join the frames table. Use load_dataset(repo, "transitions", split="all") and load_dataset(repo, "frames", split="assets"). Session record_json uses the collector tagged-tree codec and retains the complete original monitoring episode, including Run, training seed, Checkpoint, evaluation, episode seeds, recording contract, R2 hashes and start facts. Transition record_json retains original facts and the entire original monitoring_record. native_action_json is the executed provider action index, as in the reference collector; the internal emulator encoding is retained in monitoring_record.native_action and the session action contract. Unavailable separately recorded task reward/boundaries, temperature and elapsed native frames are null. Brick annotations use the reference breakout-bricks-v1 detector, including quality/initial-layout flags. No suspect frames are filtered or corrected. FirstWall episode boundaries are unchanged. R2 remains canonical. Earlier PNG export files and immutable HF revisions remain available for provenance. This publication replaces the current dataset view, not the original source recordings.

Balanced train/validation/test splits

Frozen 80/10/10 grouped split: manifest.

All recordings of the same environment seed stay in one split. Environment and policy seeds are disjoint; every one of the 20 checkpoints contributes 40/5/5 episodes. The metadata-only search prioritizes normalized brick progress, then return, and also balances episode length, progress bins, and success counts. No model performance was used.

Split	Episodes	Transitions	Mean normalized bricks destroyed	Mean return
test	100	962313	0.43782407	100.5520
train	800	7665223	0.43738426	100.4795
validation	100	946274	0.43754630	100.4890

Normalized progress uses the recorded denominator of 216. All transitions, actions, rewards, boundaries, annotations, and RGB assets are preserved. These counts precede any downstream quality or life-loss filtering. Frame assets are shared: fit only on frame IDs referenced by training transitions. This measures held-out-seed prediction within one training run, not unseen-policy or independent-run generalization. Keep test fixed and do not use it for model selection.

from datasets import load_dataset
train = load_dataset("tsilva/gradlab-breakout-c6d579da", "transitions", split="train")
validation = load_dataset("tsilva/gradlab-breakout-c6d579da", "transitions", split="validation")
test = load_dataset("tsilva/gradlab-breakout-c6d579da", "transitions", split="test")