aaung/laya-tetris-decisions

Laya Tetris decisions

Synthetic "where should this piece go?" decisions for fine-tuning a Laya typed-decision model. Each row is a text state with a table of up to eight candidate placements labelled A to H, and the preference of a heuristic teacher over those letters.

Code that makes and uses it: https://github.com/arkaung/laya-tetris Write-up: https://arkaung.github.io/laya-tetris/

What a row looks like

{
 "state": "Tetris. Piece L, next IJO, hold none. Stack height 8, holes 2, bumpiness 15. Options:\nA: clears 0, holes 2, bumpiness 15, height 8\nB: clears 0, holes 2, bumpiness 15, height 8\nC: clears 0, holes 2, bumpiness 21, height 8\nD: clears 0, holes 3, bumpiness 15, height 8\nE: clears 0, holes 2, bumpiness 13, height 8\nF: clears 0, holes 3, bumpiness 13, height 8\nG: clears 0, holes 3, bumpiness 17, height 10\nH: clears 1, holes 2, bumpiness 15, height 7",
 "answers": {
  "placement": {
   "A": 0.0,
   "B": 0.0,
   "C": 0.0,
   "D": 0.0,
   "E": 0.0,
   "F": 0.0,
   "G": 0.0,
   "H": 1.0
  }
 },
 "split": "train"
}

• state: the model input. One header line (piece, next three pieces, hold, stack height, holes, bumpiness), then one line per option: X: clears <lines>, holes <total holes after the move>, bumpiness <surface roughness>, height <max column height>.
• answers.placement: the teacher's probability for each shown letter. Letters not shown (when fewer than 8 options exist) are absent.
• split: train, val or test. Whole games go to one split, never single rows.
• questions.json (also in this repo): the question object every row is asked with. It must stay identical at training and inference time. Its instructions text and option descriptions are part of the model's input.

How it was made

1. Play many short Tetris games (60 moves each, seeded). In each game the player follows a greedy heuristic, but with a per-game exploration rate drawn from {0, 0.1, 0.3, 0.6} it plays a random one of the top options instead, so the boards range from tidy to messy.
2. At each move, list every reachable landing spot (rotation and column), compute its features, keep the best eight by the teacher score, and shuffle them so a letter carries no meaning.
3. The teacher score is 0.76 * lines_cleared - 0.51 * total_stack_height / 10 - 0.36 * holes - 0.18 * bumpiness / 4. Labels are a softmax of the scores at temperature 0.1: a clear winner gets about 1.0, near-ties share the probability. Identical option lines necessarily get equal probability.
4. Assign each game to train, validation or test (80 / 10 / 10).

Reproduce: python make_tetris_dataset.py --n 8000 --seed 0 --out tetris_data rows per split: {'train': 6169, 'validation': 911, 'test': 920}

Intended use and limits

• Fine-tuning small decision models (Laya via LayaStudio, or any classifier over letters) to read a feature table and choose. It teaches table reading, not Tetris vision: the board itself is not in the state.
• The labels are one fixed heuristic's preferences, not optimal play. A model trained on this can match the heuristic and can't beat it.
• The state shows the maximum column height, but the teacher's score uses total stack height, so labels depend slightly on something the model can't see. In practice this rarely changes the best option.
• Synthetic data from one generator and seed family. No personal data.