joelmontavon/fhir4px-model-webllm
Upload libs/fhir4px-q4bf16_1-cycle7-webgpu.wasm with huggingface_hub
Upload folder using huggingface_hub
Upload libs/fhir4px-q4f32_1-cycle7-webgpu.wasm with huggingface_hub
Upload folder using huggingface_hub
Upload libs/fhir4px-q4f16_1-cycle7-webgpu.wasm with huggingface_hub
Upload folder using huggingface_hub
Publish 3B model (24/24 regression pass, q4f16_1, ~1.7GB)
Publish 3B model (24/24 regression pass, q4f16_1, ~1.7GB)
Clean up test variants (q0bf16 broken on WebGPU, q4f32_1 untested)
Revert to prompt_fix model (19/19 regression, correct HbA1c association, varied naming). narrow_menu_fix caused degenerate outputs in WebLLM.
Publish q4f32_1 variant (4-bit weights, fp32 activations) for activation precision isolation test
Publish q4f32_1 variant (4-bit weights, fp32 activations) for activation precision isolation test
Publish q0bf16 (unquantized bf16) variant for quantization hypothesis test
Publish q0bf16 (unquantized bf16) variant for quantization hypothesis test
Add prompts.json: single source of truth for task system prompts and output shapes
Update model: narrow-menu abstention training (23/24 regression, 97% Synthea clean, over-association 7% -> 3%)
Regenerate lookup.json from full medterm benchmark (no caps, 669k entries, 141 MB)
Recompile WASM with local emsdk (fix LLVM 19/20 bitcode mismatch)
Update weights for structured prompt fix (19/19 regression pass, 98-100% held-out F1)
Upload structured app association WebGPU WASM
Upload structured app association WebLLM weights
Upload model artifacts
Upload model artifacts
Upload model artifacts
initial commit