goldenfox/decider-2b-onnx

decider 2B decision model (ONNX, WebGPU)

ONNX export of Mapika/decider-2b at revision 533964dae8be954c5b5e19fa4948e48408094c1e, for in-browser inference with ONNX Runtime Web (WebGPU execution provider, onnxruntime-web 1.30).

Conversion

• onnxruntime-genai 0.17.1 model builder, -p int8 -e webgpu, extra options prune_lm_head=true exclude_mtp=true (RTN int8 weights; token embedding kept in fp16).
• Text decoder only; max_position_embeddings set to 8192 before export (rotary cache size).
• The graph returns logits for the last input position only; recurrent (Gated DeltaNet) and KV states are exposed as past.* / present.* inputs and outputs, so a shared prompt prefix can be computed once and reused.
• External data split into files of at most 1,800,000,000 bytes.

Prompt format

Same as the source model's decider/prompt.py plain state-first layout: Context: / Question: / Options: / Answer: (, read at the last position over the label tokens and divided by the temperature in decider_config.json.

Measured agreement

9/9 top choices identical to the source model in fp32 (transformers) on 9 game prompts; mean total-variation distance 0.011 (max 0.020). Measured 2026-09-30 with onnxruntime 1.30 CPU EP.

License

Apache License 2.0, as the source model Mapika/decider-2b; see LICENSE.