LFM3-VLD-2B ONNX (WebGPU + Decision Probe)

WebGPU-ready ONNX bundle for webbrain-one/lfm3-vld-2b, with decision-trained d1-2B decoder weights, restored LFM3-VL-2B language head, and /model/layers.24/final_norm_layernorm/output_0 exposed as decision_hidden ([batch, seq, 2048]).

Technical Specifications

Backbone: LiquidAI d1-2B decision-trained hybrid conv/attention layers.
Language Head: Restored original uncorrupted LFM3-VL-2B language head (lm_head.weight).
Decision Output: Direct access to decision_hidden vector alongside standard logits.
Runtime Probes: Zero-dependency standalone JavaScript linear decision probes (lfm3-vl-decision-probe.js and lfm3-vl-decision-probe-cs.js).

Included Files

onnx/decoder_q8f16.onnx + onnx/decoder_q8f16.onnx_data (~1.9 GB)
onnx/decoder_q4f16.onnx + onnx/decoder_q4f16.onnx_data (~1.3 GB)
onnx/vision_encoder_fp16.onnx + onnx/vision_encoder_fp16.onnx_data (~362 MB)
runtime/lfm3-vl-decision-probe.js: Standard decision probe (91.3% categorical accuracy, 24/24 completions found, 0 false completions).
runtime/lfm3-vl-decision-probe-cs.js: Cost-sensitive decision probe with calibrated margin (1.071).
validation/probe.json, validation/probe-cs.json, validation/summary.json, validation/ort-parity.json.

Benchmark Performance (`final_test`, 126 browser decision tasks)

Evaluator	Accuracy	Completions Found	False Completions	Unknown Recognized
Decision Probe (`probe.js`)	115/126 (91.3%)	24/24 (100%)	0/48 (0%)	24/24 (100%)
Zero-shot verify (`verify`)	85/126 (67.5%)	22/24 (91.7%)	0/48 (0%)	0/24 (0%)

Companion Repositories

webbrain-one/lfm3-vld-2b: Full BF16 PyTorch checkpoint with untied embeddings.
webbrain-one/lfm3-vld-2b-gguf: GGUF quantizations (BF16, Q8_0, Q4_K_M, mmproj-F16) for llama.cpp.