fredericowieser/arc-agi-3-wm-traces

ARC-AGI-3 World Model Traces

This dataset contains ARC-AGI-3 transition traces in the same parquet schema used by HHazard/arc-agi-3.

Each row is one environment transition:

state, game_id, level_id, action_id, action_args, next_state,
level_done, frame_idx, origin, transformation, player

state and next_state are 64x64 ARC grids stored as nested integer arrays. action_id is the ARC-AGI-3 action kind; click actions use action_args.x and action_args.y.

Splits

Loading

from datasets import load_dataset

ds = load_dataset("fredericowieser/arc-agi-3-wm-traces", split="rollouts_positive")
print(ds[0])

Or read the parquet files directly with pandas, Polars, DuckDB, or PyArrow.

Provenance

The upload was staged from data/parquet/all_hhazard_format in the noahs-arc-wm cluster repo. The original sources were normalized to the HHazard parquet schema and validated before upload. The full consolidation manifest is included as manifest.json; source-file provenance is included as files.jsonl.

Source buckets:

• —hhazard: the original public ARC-AGI-3 human trace set.
• —prior_games: prior-game traces collected for early training.
• —rollouts: scripted/random rollout traces.
• —rollouts_positive: mined positive/terminal rollout traces.
• —rollouts_novelty_gen: novelty/frontier/l2-attack generated traces.
• —arc_static: static ARC-style transition rows converted to the same schema.

Caveats

This is a research trace dataset, not a solved simulator. It contains mixed-origin traces and should be split carefully to avoid leakage when evaluating ARC-AGI-3 generalization. Treat held-out public games and generated train-game rollouts as distinct evaluation regimes.

License

License is marked as other because the rows combine ARC-AGI-3 traces, derived/generated rollouts, and static ARC-style data. Check upstream terms before redistribution or commercial use.