Technical report: Read paper · PDF · BibTeX · All models
Collection: Spot Snake models
A Laya typed-decision model post-trained on the room-then-heading teacher for Spot's Biscuit Snake. The starting checkpoint was convaiinnovations/laya-multilingual.
| Held-out slice | Rows | Base | Post-trained | Heading baseline |
|---|---|---|---|---|
| Legacy test (English) | 1,708 | 53.57% | 100.00% | 99.94% |
| Challenge (combined) | 1,600 | 46.81% | 99.94% | 62.75% |
| Challenge (English) | 800 | 46.88% | 100.00% | 62.75% |
| Challenge (Dutch) | 800 | 46.75% | 99.88% | 62.75% |
| Challenge (heading loses, both languages) | 400 | 18.00% | 99.75% | 0.00% |
Results measure teacher agreement. The English and Dutch challenge cases use the same boards.
Choice temperature fitting returned approximately 1.0. The held-out challenge's ten-bin expected calibration error was 0.0483. See spot-snake-test.json for the full breakdown.
Game score and win rate require separate rollouts. Keep the game's legality checks in the move selection path.
The task data cover English and Dutch.
Load with laya.load() and pass the current state and a choice question containing the legal options.
See the dataset for example inputs.
onnx/model_fp16.onnx (+ model_fp16.onnx.data, 644 MB) is the whole model as one fp16 graph for
ONNX Runtime Web on WebGPU. It is what
Spot's Biscuit Snake runs for AI plays.
tokenizer.json and tokenizer_config.json sit at the root as well, for the browser.
input_ids, attention_mask [1, s] int64, marker_pos [1, k] int64 (the <mask> before each option),
marker_mask [1, k] bool, qtype [1] int64 (0 = choice). Output: logits [1, k], one per option.laya.common.build_sequence does, with max_len 1024 and head_max_len 256.onnx-check.json.conversion/, adapted from onnx-community/laya-ONNX
(Apache-2.0).The encoder and decision head were fine-tuned for three epochs using Laya's RLCD trainer on an Apple M4. The run used 5,600 training rows and reserved 400 training-split rows for temperature fitting. The scores above come from the separate test splits.
Install the runtime with pip install laya==0.4.26.
import json
import laya
from huggingface_hub import hf_hub_download
data = hf_hub_download(
"Lucasgitguy-co/spot-snake-decisions-v0.2",
"challenge_test.jsonl",
repo_type="dataset",
)
with open(data, encoding="utf-8") as stream:
example = json.loads(next(stream))
model = laya.load("Lucasgitguy-co/spot-snake-laya-multilingual-v0.1")
answer = model.predict(example["state"], example["questions"], min_confidence=0)
print(answer["answers"]["move"]["choice"])
model.safetensors 644 MBonnx/model_fp16.onnx 3 MBconvaiinnovations/laya-multilingual