FluidInference/decision-1.0-lex-snake-coreml

Decision 1.0 Lex · Snake (Core ML)

A Snake-playing variant of Decision-1.0-Lex-0.6B for on-device use. Its Choice path was fine-tuned on heuristic Snake decisions; the Noul and Score paths are unchanged from Lex. Layout and runtime match FluidInference/decision-1.0-lex-coreml: only coreml/choice.mlpackage differs, so it drops in wherever that repository is used, including FluidUse (decision-1.0-lex-snake).

Path	Tokens	Candidates	Precision
coreml/choice.mlpackage (fine-tuned)	128	3	FP16
coreml/noul.mlpackage, coreml/score.mlpackage (+ -embedding-w8 variants)	128	2 / 3	FP16 / embedding W8 (identical to Decision-1.0-Lex)

Snake

FluidUse SnakeCheck, 10×10 board, seeds 1–10. Each turn the model picks one of three non-reversing moves, each labeled with its consequence (for example left: closer to food, safe or up: crash).

	Decision 1.0 Lex	This model	Rule-based heuristic
Food, 300-move cap	112	293	305
Food, no move cap	—	345	344
Moves per game (cap)	127	292	298
Games lost to an avoidable crash	3	0	0
Median decision, M5 Pro	4.1 ms	3.9 ms	—

It relies on the move labels: with bare directions (up / left / right) it scores like the base model.

Other tasks

Same harness for both models; these are relative comparisons, not leaderboard scores.

Task	Decision 1.0 Lex	This model
Typed-decisions TEST, Choice (600)	72.0 %	71.7 %
ARC-Easy (1,000)	46.3 %	43.4 %
HellaSwag (1,000)	36.8 %	37.7 %
BANKING77 (1,000, 77 intents)	27.0 %	27.9 %
CLINC150 (1,000, 151 intents)	28.1 %	15.3 %

Use Decision-1.0-Lex for general decisions, especially routing across many options: this model leans harder on a single default answer when given long option lists.

Parity

The Core ML Choice package matches the fine-tuned native model on 200 held-out Snake decisions (max probability error 6.6e-4, every choice identical).

License

The Decision contributions are Apache 2.0. The inherited tokenizer has additional Gemma-origin terms. See LICENSE, NOTICE, LICENSING_STATUS.md, DISTRIBUTION_TERMS.md and LICENSES/, unchanged from the base model.