Beko2210/statim-decide-multilingual-base-actions

Statim Decide: actions adapter

A 13 MB LoRA adapter (rank 16) for statim-decide-multilingual-base 0.10.0 that teaches it to pick moves: a small game engine describes every legal move in one sentence, Statim reads the sentences with a one-line strategy as one choice question and returns a probability for every move. One forward pass per move, no search, no LLM. Watch it play, or let it play live.

Measured on 50 games it never saw

Evaluation seeds 1–50 for every game and player; the training games come only from seeds 100,000 and up. Mean over the 50 games; every player faces the same seeds. f32 runs merge the adapter on a GPU; q8_0 runs apply it at run time on the CPU, as the demo Space does.

Game, score	random moves	Statim, no adapter	Statim + this adapter (f32)	Statim + this adapter, q8_0 (the demo Space)	engine heuristic (teacher)
Snake, food eaten	0.18	0.10*	25.4	23.9*	22.9
Othello, disc margin	-11.8 (7 of 50 won)	-14.7 (9 of 50 won)	18.0 (44 of 50 won)	15.4 (40 of 50 won)	24.4 (47 of 50 won)
2048, score	1,132	2,010	3,401	3,840	4,721
Tetris, lines cleared	0.0	0.04	78.9	72.0	988*

* Some games of this player hit the 5,000-move cap, so its mean is a lower bound.

Paired per seed against the same games (mean difference, 95 % bootstrap interval, wins / losses, two-sided sign test):

Game	against	mean difference	wins / losses	p
Snake	without the adapter	+25.3 [+23.4, +27.1]	50 / 0	1.8e-15
Snake	the teacher	+2.5 [+0.2, +4.9]	30 / 18	0.11
Othello	without the adapter	+32.7 [+25.5, +39.7]	46 / 4	4.5e-10
Othello	the teacher	-6.4 [-11.6, -1.3]	17 / 27	0.17
2048	without the adapter	+1,391.0 [+879.1, +1,938.6]	39 / 11	9e-05
2048	the teacher	-1,319.8 [-2,278.0, -445.5]	17 / 33	0.033
Tetris	without the adapter	+78.8 [+58.8, +100.1]	50 / 0	1.8e-15
Tetris	the teacher	-909.3 [-1,101.1, -723.8]	2 / 48	2.3e-12

Without the adapter the model follows the wording of the move names, not the sentences. With it, Snake and Othello reach the teacher; 2048 and Tetris stay below it. The adapter only acts on requests that name it ("adapter": "actions"); every other request is served by the unchanged base model.

Use

statim serve -m multilingual=statim-decide-multilingual-base-q8_0.gguf \
    --adapter multilingual:actions=statim-decide-multilingual-base-actions.lora.gguf

Requests: demos/actions/prompting.py (build_request(game, labels="moves")) in the repository, or site/actions/engines.js in the browser.

Files

File	Size
statim-decide-multilingual-base-actions.lora.gguf	13.5 MB
peft/adapter_model.safetensors	13.5 MB

Checksums in SHA256SUMS. The adapter is bound to the 0.10.0 base (its checkpoint fingerprint); Statim refuses to load it onto another model.

Training

tools/finetune/train_lora.py --rows on 119,952 decisions of the engine heuristics (2048 60,000, othello 11,274, snake 17,117, tetris 31,561), generated by demos/actions/make_data.py: soft targets from the heuristic's move scores, 15 % random exploration moves, shuffled option order and 50 % letter labels. LoRA r 16, alpha 32, best dev agreement with the teacher 0.749 (before 0.308). The data is generated by our own code; no external dataset is used.

Licence

PolyForm Noncommercial 1.0.0 only, like its base model 0.10.0, whose training data includes non-commercial sources (details). Text in LICENSE-MODEL.md. The Statim engine is Apache-2.0.