makemake/5k-balatro-games

5,000 Balatro games

5,000 complete games, 1,126,644 decision records, generated by the V68 heuristic in the Pylatro headless simulator. There are 3,728 wins and 1,272 losses (74.56% wins). Both outcomes are included.

The dataset is stored directly as 20 Parquet files, each containing 250 complete games. No ZIP extraction, pickle loading, simulator installation, or custom Hugging Face loading script is required. A row is one decision, not one game. seed identifies the game, and step gives its zero-based decision index.

Generation

• Blue Deck (b_blue), White Stake (1), victory through the Ante 8 boss.
• Selected V68 policy, using search shops and scaler growth; blind rollout disabled.
• The teacher has full simulator state and a current score/RNG oracle. Its information is richer than the student's saved observations.
• Discount factor 0.997; shaped reward configuration is recorded in full in metadata/schema.json, with reward model version and fingerprint.
• Tokenizer version 13, semantics v13_phase_risk_economy_card_state.
• Generated September 10–11, 2026. The exact seed list and source/binary hashes are in metadata/generation.json.

Generation originally targeted 100,000 games across parallel workers, then stopped at the user's requested 5,000-game checkpoint. This release contains exactly those 5,000 selected complete games, sorted by seed; it is not a random sample of all possible seeds. Seeds begin at 2,000,000 across disjoint worker ranges. One worker shard ended early, so those ranges have unequal coverage. The reserved evaluation seeds (10,000–10,999 and the separate reserved random panel) were excluded. No game was selected by its win/loss outcome.

The observed win rate describes this dataset; it is not an independent estimate of general heuristic strength. One deck/stake setting and one teacher are represented. No claim of balanced hand strategies is made.

Decision fields

Field	Meaning
seed, step	Game identifier and decision order within the game.
obs	Tokenized observation, legal-action mask, scalar features, and played-hand history.
action	Teacher's flat action ID; see metadata/actions.json.
reward	Immediate shaped reward from taking this action.
return_target	Discounted sum of rewards from this decision to the end of the game.
won, max_ante	Episode outcome and maximum recorded Ante. A winning cash-out can advance the terminal state to Ante 9.
terminal_outcome_target, terminal_outcome_mask	Outcome class: 0–7 means loss in Ante 1–8; 8 means victory. The validity mask is 1 for every published record.
win_ante, deck_key, stake	Game configuration.
tokenizer_version	Observation schema version.
heuristic_*	Shop-policy, scaler-growth, and blind-rollout settings.

Array shapes and exact numeric dtypes are in metadata/schema.json. Token IDs and action IDs belong to this Pylatro schema, rather than universal Balatro IDs. metadata/vocabulary.json, metadata/constants.json, and metadata/actions.json provide the recorded vocabulary, token layout constants, and action decoding.

Validation and splits

Every game was checked for legal teacher actions, finite observations/returns, valid terminal labels, and successful loading through the original training loader and collator. The original archive passed a full CRC check. During conversion, every Parquet row group was compared for exact Arrow value equality against its original episode, preserving numeric dtypes and array shapes. metadata/summary.json lists record counts, byte sizes, and SHA256 hashes for every Parquet file; metadata/games.jsonl lists per-game outcomes and lengths.

Only a train split is published. For validation or test splits, divide by seed, keeping all decisions from a game together. Randomly splitting individual decisions would leak trajectories across splits. These are bootstrap imitation-learning demonstrations, not a held-out benchmark.

License

MIT, retaining the license selected for this dataset repository by its owner.