onnx-community/strands-decider-2B-hobson-v19-ONNX

strands-decider-2B (hobson v19), ONNX

ONNX conversion of StrandsAgents/strands-decider-2B-hobson-v19, Strands Labs' decision model: give it a state and typed questions (choice, score, noul), and it returns a calibrated probability for every option, in one forward pass per question. For the model, its training data, its evaluation and its limits, see the original card and the strands-decider repository.

It runs in the browser on WebGPU with open-jev (model: "strands-decider-2b"), or with ONNX Runtime anywhere.

How it was made

strands-decider-2B is Qwen3.5-2B-Base (at the revision the adapter was trained on) with a rank-16 LoRA adapter and a pointer head of about a million parameters on the final hidden states. This repo:

• merges the LoRA into the base weights (W + alpha/r · B·A, alpha/r = 2);
• puts them into the graphs of onnx-community/Qwen3.5-2B-ONNX-OPT (fused LinearAttention / CausalConvWithState), mapped by name with the transforms found by value for the same graphs (transpose, 1 + weight for the zero-centred RMSNorm, -exp(A_log)); every swapped weight has cosine similarity ≥ 0.993 to the instruct weight it replaces;
• drops the LM head (the decider never reads logits) and exposes the final hidden states;
• fuses embeddings, decoder (with empty caches and position ids built in the graph) and the pointer head into one graph:

input_ids [B, L], attention_mask [B, L], answer_pos [B], option_pos [B, K]  ->  logits [B, K]

logits are before temperature: divide by decider.temperature_by_kind[type] from config.json (noul 0.911, choice 0.734, score 1.328) and take a softmax over the question's options.

Scripts and the reference cases are in conversion/.

Files

dtype (Transformers.js)	File	Weights	Size
q8 (default)	onnx/model_quantized.onnx	decoder 8-bit, block 32 (fp16 activations); embeddings 4-bit	1.80 GB
q4f16	onnx/model_q4f16.onnx	decoder and embeddings 4-bit, block 32	1.09 GB

Both need WebGPU with shader-f16 in the browser (ONNX Runtime's CPU provider runs them too). q8 here is 8-bit block quantization with ONNX Runtime's own kernel (MatMulNBits, bits=8), not the usual dynamic int8 export; it uses the q8 slot because Transformers.js has no q8f16.

Checks

Against the upstream engine (strands_decider.infer, PyTorch, CPU fp32), largest difference in any option's probability and how often the top answer changed:

Runtime	11 questions on 6 short states	40 questions on 8 typed-decisions states
fp32 ONNX (ONNX Runtime, CPU)	0.0009, 0 changed	0.0013 (10 questions), 0 changed
q8, WebGPU (Chrome, M3 Pro)	0.016, 0 changed	0.026, 0 changed
q4f16, WebGPU	0.072, 0 changed	0.146, 5 changed

q4f16 moves near-ties on long states; prefer q8 unless the download size matters more.

The open-jev integration was also checked in Node against the Python run of the same graph: identical probabilities (prompt rendering, tokenization, option positions and temperatures match prompting.py and infer.py).

Speed and accuracy in the browser

On a MacBook Pro (M3 Pro, Chrome, WebGPU): about 130 ms for one short question; questions on one state run as a batch. On the 400 states × 5 questions of typed-decisions (test split, top answer = annotator label, as in the open-jev demo's benchmark):

dtype	choice	score	noul	overall	median per state
q4f16	334/600	433/800	401/600	58.4%	2.2 s
q8	337/600	458/800	392/600	59.4%	2.2 s

The model was not trained on typed-decisions. Its confidence holds up there too: answers at 0.9 or more are right 94% of the time (q8).

License

Apache-2.0, as the original model and Qwen3.5-2B-Base.