caries-medgemma-gguf

caries-medgemma (GGUF)

GGUF-quantized weights of a LoRA fine-tune of google/medgemma-1.5-4b-it on a patient-safe intraoral caries dataset. Intended for use with Ollama as the per-crop verdict model in a YOLO + MedGemma two-stage pipeline.

Companion repo (code, dataset split, evaluation): https://github.com/Eeman1113/dental_study

Files

• medgemma_caries-Q4_K_M.gguf (2.3 GB) — quantized text model
• medgemma_caries-F16.gguf (7.2 GB) — full-precision text model
• mmproj-medgemma_caries-F16.gguf (812 MB) — vision projector (required for image input)

Usage with Ollama

huggingface-cli download sdvzdfgfngdfgh/caries-medgemma-gguf \
  medgemma_caries-Q4_K_M.gguf mmproj-medgemma_caries-F16.gguf \
  --local-dir ./gguf

cat > Modelfile <<EOF
FROM ./gguf/medgemma_caries-Q4_K_M.gguf
FROM ./gguf/mmproj-medgemma_caries-F16.gguf
TEMPLATE \"\"\"{{- range \$i, \$_ := .Messages }}
{{- if eq .Role \"user\" }}<start_of_turn>user
{{ .Content }}<end_of_turn>
{{ else if eq .Role \"assistant\" }}<start_of_turn>model
{{ .Content }}<end_of_turn>
{{ end }}
{{- end }}<start_of_turn>model
\"\"\"
PARAMETER stop <end_of_turn>
PARAMETER temperature 0.1
EOF

ollama create caries-medgemma -f Modelfile

Caveats

Whole-image detection is collapsed (always emits the negative template). Only useful as the per-crop verdict model in the two-stage pipeline. See the eval report in the companion GitHub repo for numbers.

Not for clinical use. See MedGemma's HAI-DEF terms.