guilindev/pacman-decision-oracle-v1

Pac-Man Decision Oracle — v0.1 training data

Fresh synthetic game states and soft decision labels used to fine-tune guilindev/pacman-decision-qwen3-0.6b. The pilot contains 4,096 training examples and 512 validation examples. There are no customer records, human annotations or game screenshots.

How it was built

The deterministic open-source decision-pacman engine, pinned at 593ed1f59f4f97ad0bda7287ff304e9667360089, supplies current-state observations. For each legal move, a search teacher simulates five seconds of play with greedy continuation and scores the outcome. Its values are converted to soft targets with temperature 50. This is search-teacher distillation; no LLM teacher or Jev API supplied labels.

Three collecting policies generate diverse states: oracle, greedy and seeded random. Training collectors use seeds 10000–10014, 11000–11019, and 12000–12039. Validation collectors use 900–902, 910–913, and 920–927. Games run at speed 1, with a lockstep clock and a maximum of 180 simulated seconds.

The raw pools contain 42,555 training rows and 10,048 validation rows. Exact duplicate prompts are removed within and across splits, reserving validation first. This leaves 25,205 training and 5,935 validation rows. After independently shuffling each pool with Python random seed 42, this pilot selects the first 4,096 / 512. data_manifest.json records the released file hashes and actual seed membership.

The student input includes the current paths, nearby ghosts, food, power pellets, lives and legal options. Future rollouts, oracle values and labels are excluded from the input messages. The soft targets remain available as supervision.

Fields

Field	Meaning
messages	System and user messages; score the first answer position with thinking disabled
keys	Legal moves in the order assigned to A, B, C, D
target	Softmax over the search teacher's move values; sums to one
label	Teacher's selected move
seed, tick, player	Reproducible collection provenance
prompt_sha256	Exact-input deduplication signature

The published model normalizes candidate scores only over the listed legal options. Neither a target value nor a model score is a calibrated probability of winning.

Reproduce and use

from datasets import load_dataset
data = load_dataset('guilindev/pacman-decision-oracle-v1')
print(data['train'][0]['keys'])

The model repository includes training code and the pilot evaluation protocol. The 100–109 seed range is reserved for independent gameplay; no such seed is present in either released split. Seeds vary game trajectories on one fixed maze, so a seed holdout is not an unseen-map test.

Attribution

Game engine, feature encoder, search oracle and prompt format originate from grapeot's MIT-licensed decision-pacman. Its license is included as LICENSE.upstream. Data collection, cross-split deduplication, this data release and the new fine-tuning run were performed for this project.

The released synthetic data is Apache-2.0. It is intended for research and teaching about specialized decision models. It does not evaluate general judgment, real-world visual understanding, or industrial inspection.