trinacratech/snake-trap-onnx

snake-trap (ONNX)

Laya laya-multilingual fine-tuned to play Snake from a text description of the board, taught by a hand-written rule planner. Superseded by trinacratech/snake-rl-room-onnx, which reads the same input but learned from a reinforcement-learning teacher (102 vs 85 points). Kept for comparison: it is selectable in the "hard" mode of laya-pi.

file	what
model-fp32.onnx	the model, 1.29 GB, fp32, single file
rl_agent_config.json	encoder name, max_len, calibration temperatures
tokenizer/	tokenizer
encoder/config.json, snake_report.json	encoder config; evaluation report

Input

Identical to snake-rl-room: four lines of head-relative board facts and a neutral move choice (UP/DOWN/LEFT/RIGHT). The only difference is the optional trap question:

"If the head moves one cell <dir>, is the snake trapped: dead, or cut off from its own tail?"

Build prompts with laya-pi's snakeweb/snake_state.py; other wording plays worse.

Training

• Base: convaiinnovations/laya-multilingual (encoder jhu-clsp/mmBERT-base).
• Teacher: a rule planner (breadth-first search over the board) giving the best move(s) and a "tail still reachable" trap label.
• 55,134 training items, 4 epochs, 87 minutes on one GPU.

Evaluation

2,000 held-out boards and 20 full games (24 × 16, seeds 50000–50019), real game engine, no planner help.

	this model	laya-multilingual zero-shot
mean score, plain	85.2 (median 91, max 107)	0.05
mean score, trap shield	90.8 (median 94, max 119)	—
any best move chosen	96.7%	52.8%
fatal move chosen	0.0%	11%

One full game on a Raspberry Pi 4 (ONNX Runtime fp32, no shield): 108 points, length 114, 2,273 moves, ~3 s per move.

Limitations

24 × 16 boards only, exact prompt only; greedy on long snakes. For demonstration and teaching.

Credits

Laya by Convai Innovations (Apache-2.0). Snake rules from laya-mlx (Apache-2.0).