tsilva/gradlab-breakout-6127e81d

Breakout checkpoint trajectories

500 episodes and 2,059,758 transitions, split 80/10/10 by whole evaluation-seed groups. The same environment or policy seed never appears in different splits, including across checkpoints. Every checkpoint contributes 40 train, 5 validation and 5 test episodes.

Start here

Browse the Breakout environment collection.

This dataset contains full-HUD Breakout trajectories from Checkpoint Monitoring of Run gradlab-6127e81defbc9774d9a57a0ff55e2f6e. It is observational data from saved Checkpoints, not Acceptance or a released Policy. This page currently presents an explicit train/validation/test view; the original unsplit snapshot remains at the immutable revision linked below.

Load transitions for actions and frame IDs, frames for shared lossless RGB images, episodes for boundaries and split membership, and sessions for Run and Checkpoint provenance. The code example below loads the transition splits.

This Run differs from the larger full-HUD Checkpoint dataset. The separate collector dataset masks the HUD.

Split	Episodes	Transitions	Mean bricks	Mean shaped return	Mean length	First-wall successes
train	400	1647968	85.60	87.87	4119.92	55
validation	50	206044	85.16	87.48	4120.88	7
test	50	205746	85.46	87.79	4114.92	7

Splits balance the brick-progress distribution first and also return and episode length. Full method, seed assignments, distribution statistics, validation thresholds and source revision: split manifest.

from datasets import load_dataset
train = load_dataset("tsilva/gradlab-breakout-6127e81d", "transitions", split="train", streaming=True)
validation = load_dataset("tsilva/gradlab-breakout-6127e81d", "transitions", split="validation", streaming=True)
test = load_dataset("tsilva/gradlab-breakout-6127e81d", "transitions", split="test", streaming=True)

Frames remain deduplicated lossless WebP with the full unmasked HUD, and the trajectory schemas and brick annotations are unchanged. frames uses the shared assets split; sessions uses metadata. Join frame IDs from each split's transitions. Fit an encoder or preprocessing only on training-referenced frames, not the entire frame pool. Identical states can naturally recur across episodes; shared RGB IDs are counted in the manifest. This is held-out-seed evaluation within one training run, not unseen-policy generalization. Use validation for model selection and reserve the frozen test split for final evaluation. Unavailable source facts remain null; native_action_json is the submitted provider action index and record_json preserves the original internal encoding and transition evidence. Session records preserve original episode provenance. Their original split assignment was null; the episode table and split manifest define the new assignment.

The unsplit dataset remains available at its immutable revision. Original R2 recordings and earlier HF files are preserved.

Dataset schema

Trajectory schema 1, contract fingerprint 3af3a3bf898395756eb357517276785695898d3b6f4cc584b59e971dc4d60bd6. Schema definition; migration receipt.

Migrated from revision 79827bb. All table values, image bytes, episode assignments and split statistics were preserved and verified. Every Parquet shard carries the schema version, fingerprint and table identity.