FieldMedic Gemma 4 E2B LoRA

Gemma 4 E2B IT fine-tuned for offline medical triage. Outputs structured JSON for the FieldMedic Pocket Android app.

This is the LoRA adapter for the FieldMedic Pocket project, submitted to the Gemma 4 Good Hackathon (Kaggle × Google DeepMind, May 2026).

What this model does

Takes a natural-language description of an injury or symptom (in English, Hindi, Bengali, Tamil, or Marathi) and returns a structured JSON triage response with one of four function calls:

report_triage — severity (RED / YELLOW / GREEN) + immediate steps + do-not list + watch-for list
escalate_emergency — reason + single most important action + while-waiting steps
request_retake — instructions for clearer input
refuse_out_of_scope — redirect to appropriate resource

This model does not diagnose. The JSON schema has no diagnosis field. It triages severity and recommends actions, never names a condition.

Intended use

Designed to run on-device via LiteRT-LM on Android phones in low- or no-connectivity environments. Primary users: rural community health workers, family caregivers, volunteer disaster responders, displaced civilians.

How to use

As a LoRA adapter with Unsloth

from unsloth import FastLanguageModel
import torch

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="unsloth/gemma-4-e2b-it-bnb-4bit",
    max_seq_length=2048,
    dtype=None,
    load_in_4bit=True,
)
model = FastLanguageModel.get_peft_model(model, ...)  # see Unsloth docs

# Load this adapter on top
model.load_adapter("VishavGarg/fieldmedic-gemma-4-e2b-lora")
FastLanguageModel.for_inference(model)

# Inference
messages = [
    {"role": "system", "content": FIELDMEDIC_SYSTEM_PROMPT},
    {"role": "user", "content": "Child burned hand on hot oil, blistering visible"},
]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text=prompt, return_tensors="pt").to("cuda")
output = model.generate(**inputs, max_new_tokens=400, temperature=0.1, do_sample=False)
print(tokenizer.decode(output[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

As merged FP16 weights (for export to LiteRT-LM / .litertlm)

Coming soon — merged weights upload after the demo video is recorded.

Training details

Item	Value
Base model	`unsloth/gemma-4-e2b-it-bnb-4bit`
Framework	Unsloth + 🤗 TRL `SFTTrainer`
Dataset	420 hand-curated medical-triage examples (custom, not yet publicly released)
Epochs	3
LoRA rank	16
LoRA alpha	16
LoRA dropout	0
Target modules	`q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj`
Batch size	2 (× grad-accum 4 = effective 8)
Learning rate	2e-4
LR scheduler	linear
Mixed precision	bf16 (Ampere+) / fp16 (T4) auto-detected
Optimizer	adamw_8bit
Max seq length	2048
Training time	~45 min on T4 (free Colab)
Final training loss	TBD (filled after training completes)

Evaluation

20-case held-out test suite. Pass rate: 17/20 (85%) — see GitHub repo for the full case list.

Metrics:

Structured-output rate (valid JSON): 100%
Schema conformance: >95%
Severity correctness (RED/YELLOW/GREEN): 85%
Multilingual coverage: English, Hindi, Bengali, Tamil, Marathi

Safety

No diagnosis: schema explicitly excludes diagnosis field
Always-escalate on RED severity — one-tap dial to local emergency number
Refuses medication dosing, mental-health, dental, veterinary, food-safety queries
Visible disclaimer on every screen: "Not a doctor. Always consult a medical professional when possible."
Trained only on hand-curated triage scenarios — no PubMed scrape, no protected medical records

Limitations

Knowledge cutoff inherits from Gemma 4 base — does not know post-2024 medications/protocols
Animal-bite cases sometimes incorrectly refused as out-of-scope (3/20 failures documented)
Hindi response sometimes leaks English numerals (cosmetic, harmless)
Borderline YELLOW/RED severity drift on ambiguous cases

Citation

@misc{garg2026fieldmedic,
  title={FieldMedic Pocket: Offline Medical Triage on Gemma 4 E2B},
  author={Garg, Vishav},
  year={2026},
  publisher={Hugging Face},
  url={https://huggingface.co/VishavGarg/fieldmedic-gemma-4-e2b-lora}
}

License

Apache 2.0, inheriting from Gemma 4. See Gemma Terms of Use.

Acknowledgements

Google DeepMind — Gemma 4
Daniel & Michael Han — Unsloth
The medical-triage protocols informing the system prompt were derived from publicly available first-aid guides (Red Cross / WHO / St. John Ambulance — see 02-triage-system-prompt.md)

Submitted to the Gemma 4 Good Hackathon, May 2026.

Downloads last month: 20