Modelpublic

costinstroie/medical-gemma-270m-GGUF

sourceHugging Facegemmaupdated 9d agoView on Hugging Face

0likes95downloads

Model Card

medical-gemma-270m-GGUF

GGUF conversion of vervelo/medical-gemma-270m-mlx, a LoRA adapter (rank 16) for google/gemma-3-270m-it fine-tuned to turn medical dialogues into SOAP notes.

The MLX adapter was merged into the base weights (W + scale·(A·B)ᵀ, fp32, then stored as bf16) and converted with llama.cpp (b8733). All credit for the fine-tune goes to the original author; this repo only repackages it.

File	Size
`medical-gemma-270m-Q8_0.gguf`	279 MB
`medical-gemma-270m-Q4_K_M.gguf`	242 MB

Usage

The embedded chat template reproduces the training prompt: the user message is the dialogue, and an optional system message replaces the default instruction.

bash

llama-cli -m medical-gemma-270m-Q8_0.gguf
# paste: Doctor: What brings you in? Patient: I have had a cough and fever for three days.

Raw-prompt equivalent:

### Instruction: Convert the following medical dialogue into a structured SOAP note (Subjective, Objective, Assessment, Plan).

### Input: Dialogue: <dialogue>

### Response: SOAP Note:

Q80 is recommended; at 270M parameters Q4K_M loses noticeable quality.

Limitations

Very small model; output is often terse. Trained on synthetic dialogues. Not for clinical use — outputs may be wrong and must be reviewed by a qualified professional.

License

Subject to the Gemma Terms of Use, inherited from the base model. Training data licenses: MTS-Dialog CC BY 4.0; MedSynth has no explicit license (research use).