BoogieKn/laya-plays-google-snake

Laya Plays Google Snake — adapted checkpoint

This is the full checkpoint used by Laya Plays Google Snake, a controller for the real Google Snake browser game. It scores safe UP, RIGHT, DOWN, and LEFT options from a structured board description. It does not parse screenshots or issue keypresses by itself; the open-source controller handles those steps and filters unsafe moves.

Origin and training

The base is the multilingual checkpoint from Convai Innovations' Laya, released under Apache-2.0. This adaptation retains its mmBERT-base encoder and tokenizer. The encoder was frozen while Laya's typed-decision layers were trained by supervised imitation on balanced Snake move examples generated by the project's simulator. It was not trained end-to-end with reinforcement learning on live browser screenshots.

The bundled rl_agent_config.json records snake-balanced-imitation-frozen-encoder. It does not record the exact hyperparameters of the original training run, so retraining is not guaranteed to recreate identical weights. The project publishes its training code.

Download and run

Install uv and clone the project, then download this model into the path expected by its CLI:

git clone https://github.com/Makankn/laya-plays-google-snake.git
cd laya-plays-google-snake
uv sync --extra laya
uv run --extra laya hf download BoogieKn/laya-plays-google-snake --local-dir models/snake-laya-balanced
uv run --extra laya snake-state eval-laya --model models/snake-laya-balanced --device cuda --episodes 10 --max-steps 160 --seed 10000 --validation-samples 0

For live play, see the project README. The tested accelerated setup uses Windows and an NVIDIA GPU; --device cpu is available but has not been validated for live input timing.

This is a custom Laya checkpoint, not a standard Transformers pipeline or image-to-action model. It requires the project's snake_laya policy/runtime and should not be expected to run in the default Hugging Face inference widget.

Evaluation

On 2026-09-26, the project's local evaluator ran 10 fixed-seed simulator episodes at 160 steps each. The checkpoint averaged 12.6 apples per episode (best 15, worst 6). One episode repeated positions for 71 steps. These are simulation results, not measured live Google Snake scores; the test does not include visual tracking errors, browser focus loss, occlusion, or keyboard timing.

The model was adapted to a 17×15 classic Google Snake board and compact English move descriptions. Its output probabilities compare the offered choices; they are not calibrated probabilities of live-game survival. The surrounding controller may exclude a move or override the top prediction when safety or tracking requires it. This experimental checkpoint is not appropriate for safety-critical control.

Files, license, and attribution

model.safetensors contains the full weights; rl_agent_config.json, encoder/config.json, and tokenizer/ contain the configuration and tokenizer needed by the project loader. See LICENSE for Apache-2.0 terms and NOTICE for upstream attribution. This independent project is not affiliated with Google or Convai Innovations.