v2.1-instruct-medgemma-bf16

v2.1 Instruct MedGemma (SFT-only)

A clinical decision-support model fine-tuned from MedGemma 1.5 4B-IT on the Uganda Clinical Guidelines 2023. Conservative SFT-only checkpoint with perfect refusal compliance.

For the primary production model (SFT + DPO, best overall performance), see Crane MedGemma 1.5 IT.

This model is a clinical thinking aid. It does not provide diagnoses, prescriptions, or automated clinical decisions.

Intended Use

Primary use case: Assist frontline health workers (clinical officers, nurses) at Health Centre II–IV facilities in Uganda with clinical triage reasoning — differential diagnosis, danger sign identification, investigation ordering, referral criteria, and special population considerations.

Output formats: The model can emit structured triage assessments in three formats (XML, positional array, prose) based on the system prompt. The production app uses compact XML packets for low-latency on-device inference.

Deployment target: On-device inference on Android via quantized GGUF. This bf16 checkpoint is the full-precision reference and was used as the base for the DPO-aligned Crane MedGemma 1.5 IT.

In-Scope

• Differential diagnosis reasoning for conditions in the Uganda Clinical Guidelines
• Danger sign identification and referral triage
• Investigation recommendations
• Special population considerations (pediatric, pregnancy, HIV)
• Refusal of out-of-scope queries (treatment, dosing)

Out-of-Scope

• Treatment and dosing recommendations — the model is trained to refuse these queries and redirect to the facility's Uganda Clinical Guidelines
• Diagnostic conclusions or prescriptions
• Multi-turn conversations
• Languages other than English
• Conditions not covered by the Uganda Clinical Guidelines 2023

Model Details

Training Approach

The model was fine-tuned in two stages:

1. Decision-support SFT: QLoRA fine-tuning on 11,991 Q&A pairs covering 7 clinical categories (differential diagnosis, diagnosis, referral, investigation, danger signs, special populations, refusal). Treatment and dosing categories were excluded after evaluation showed a capacity ceiling on factual drug recall at this parameter scale.

2. Instruction-following SFT: Continued fine-tuning to teach format-switching — the model emits XML, array, or prose triage packets depending on the system prompt.

What the Model Refuses

The model is trained to refuse treatment and dosing questions with: "Treatment and dosing recommendations are outside the scope of this tool. Please refer to your facility's Uganda Clinical Guidelines."

This is a deliberate safety boundary. Drug name confusion at the 4B parameter scale made treatment responses unreliable, so the scope was narrowed to decision-support where the model performs well.

Evaluation

Evaluated across three benchmarks using Gemini as an automated evaluator in constrained comparison mode.

Ship Gate Analysis

5 of 12 ship gates passed. Content quality on held-out conditions caps at ~3.0/5 — a parameter capacity ceiling, not a training procedure failure. Retrieval-augmented generation (RAG) is expected to close this gap.

Strengths

• Perfect refusal compliance: 5.00/5 on prose refusal — never leaks treatment or dosing content
• Strong differential diagnosis: 4.38/5 on diagnosis questions
• High XML parse rate: 98.2% valid structured output
• Reasoning over recall: "Why" questions score 5.0/5; the model reasons well about clinical logic

Known Limitations

• 4B parameter capacity ceiling: Held-out content quality caps at ~3.0/5. The model has limited factual recall for conditions it hasn't seen extensively in training.
• Treatment and dosing excluded: By design. The model refuses these queries.
• Special populations weakness: 1.11/5 — the source guidelines have limited population-specific detail for many conditions.
• Array parse rate: 92.2% (below 95% gate). XML and prose formats are more reliable.
• Single-turn only: No conversational follow-up capability.

How to Use

from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "CraneAILabs/v2.1-instruct-medgemma-bf16"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype="bfloat16", device_map="auto")

messages = [
    {"role": "system", "content": "You are a clinical decision-support tool based on the Uganda Clinical Guidelines. Provide structured triage assessments. Do not provide treatment or dosing recommendations."},
    {"role": "user", "content": "A 4-year-old presents with fever 39.5C for 3 days, neck stiffness, and photophobia. What should I consider?"}
]

inputs = tokenizer.apply_chat_template(messages, return_tensors="pt", add_generation_prompt=True).to(model.device)
outputs = model.generate(inputs, max_new_tokens=512, temperature=0.3)
print(tokenizer.decode(outputs[0][inputs.shape[-1]:], skip_special_tokens=True))

Ethical Considerations

• Not a replacement for clinical judgment. All outputs are advisory. The clinician makes all decisions.
• No patient data was used in training. All training data is derived from published government clinical guidelines.
• Offline-first deployment — no patient data leaves the device.
• Scope boundaries are safety boundaries. The refusal of treatment/dosing queries is a deliberate design choice to prevent harm from drug name confusion at this model scale.

Citation

If you use this model, please cite:

@misc{crane-medgemma-v21-2026,
  title={v2.1 Instruct MedGemma: SFT-Only Clinical Decision-Support for Uganda Clinical Guidelines},
  author={Crane AI Labs},
  year={2026},
  publisher={Hugging Face},
  url={https://huggingface.co/CraneAILabs/v2.1-instruct-medgemma-bf16}
}

License

This model is subject to the MedGemma Terms of Use. Additional fine-tuning artifacts are proprietary to Crane AI Labs.