LFM3-VL-2B-VLD Multipass
LiquidAI/LFM3-VL-2B adapted with WebBrain's multipass VL-D (verify Yes/No per option + readout option code) decision protocol, distilled from LiquidAI/d1-2B on WebBrain multimodal browser workflow screenshots — without full VL/Compass fine-tuning (vision tower frozen, language backbone adapted via LoRA and merged to BF16).
Companion Repositories
webbrain-one/lfm3-vld-2b-multipass-onnx: WebGPU ONNX bundle (decoder_q8f16.onnx,decoder_q4f16.onnxwithdecision_hiddenexposed,vision_encoder_fp16.onnx, and trained linear decision probes).webbrain-one/lfm3-vld-2b-multipass-gguf: GGUF exports (BF16,Q8_0,Q4_K_M, andmmproj-F16) forllama.cpp(/v1/chat/completions+/v1/systemone).webbrain-one/lfm3-vld-2b-multipass-lora: LoRA adapter (r=32, alpha=64) and projector weights.
Results on d1 Balanced Browser Decision Benchmark (final_test, 126 Categorical Questions)
| Model / Protocol | Correct | Completions Found | False Completions | Unknown Recognized | Score MAE |
|---|---|---|---|---|---|
LFM3-VLD-2B (probe / probe-cs) |
126/126 (100.0%) | 24/24 | 0/48 | 24/24 | 9.56e-8 |
LFM3-VLD-2B (verify zero-shot) |
126/126 (100.0%) | 24/24 | 0/48 | 24/24 | 4.11e-5 |
LFM3-VLD-2B (readout) |
125/126 (99.2%) | 24/24 | 0/48 | 24/24 | 0.0189 |
Base LiquidAI/LFM3-VL-2B (verify zero-shot) |
106/126 (84.1%) | 24/24 | 0/48 | 4/24 | 0.0004 |
Base LiquidAI/LFM3-VL-2B (readout) |
102/126 (81.0%) | 24/24 | 1/48 | 0/24 | 0.1106 |