Dronam-Vision-V1 — GGUF (Q4_K_M) + fracture classifier
Dronam-Vision-V1 merged into Gemma-4-E2B-it and quantised for llama.cpp / LM Studio, plus the ONNX fracture classifier this model needs for its fracture decision (see the main card: the adapter alone does not detect fractures reliably).
Held-out results (816 FracAtlas radiographs): hybrid balanced accuracy for fracture detection 85.8% (93.0% sensitivity, 78.6% specificity), region accuracy 92.3%, view accuracy 87.7%. Full table and limitations on the main card. Drafting aid only; a clinician reviews every read.
| File | What |
|---|---|
gemma-4-e2b-it.Q4_K_M.gguf | the vision-language model (3.4 GB) |
gemma-4-e2b-it.BF16-mmproj.gguf | vision projector (1.0 GB) |
radiograph_instruction.txt | the exact instruction used in training |
radiograph.schema.json | the output schema for structured decoding |
fracture_classifier/model.onnx | fracture classifier (CPU, ~40 MB) — decides fracture_present |
fracture_classifier/config.json | image size, normalisation, decision threshold |
fracture_classifier/fracture.py | reference inference code (onnxruntime) |
ATTRIBUTION.txt | FracAtlas licence and citation |
Run
lms import -c -y --user-repo dronam/dronam-vision-v1 gemma-4-e2b-it.Q4_K_M.gguf
lms load dronam-vision-v1 --context-length 4096 -y
Call /v1/chat/completions with an image and radiograph_instruction.txt, temperature: 0, and
response_format set to radiograph.schema.json for schema-constrained output — this gives region,
view and hardware. Separately, run the image through fracture_classifier/ (onnxruntime, CPU) for the
fracture flag, and use that instead of whatever fracture_present the language model returns.
Licence: Apache-2.0.