A LoRA-fine-tune of google/gemma-4-e2b-it specialized for ASHA-Saathi — an offline, voice-first AI co-pilot for India's ~1 million ASHA frontline community-health workers, in Hindi and Marathi.

Submitted to the Gemma 4 Good Hackathon. Repo: github.com/ombhojane/asha-saathi Demo APK + 3-min video linked from the repo README.

Intended use

Decision-support for ASHA workers in the field, offline, on a low-end Android (≤4 GB RAM, Snapdragon 4-gen / Dimensity 6020 class). Specifically:

Maternal & child-health protocol Q&A (ANC/PNC, ORS, vaccinations, anemia, malnutrition)
Native function calling for dosage_calculator, vaccine_schedule, danger_sign_check, nearest_phc_referral
Out-of-scope refusal (refers up when asked about cancer, antibiotics, surgical decisions, etc.)
Danger-sign triage (IMNCI matrix)

Not for: primary clinical decision-making, replacing doctors, English-only deployments, languages outside Hindi/Marathi.

How it was trained

Base: google/gemma-4-e2b-it
Method: QLoRA (4-bit) via Unsloth, rank 64, alpha 64, all linear modules
Dataset: ombhojane/asha-instructions-hi-mr-v1 — 5–8k examples, 60% protocol Q&A / 25% function-call / 10% refusal / 5% danger-sign
Hyperparams: lr 2e-4 cosine, warmup 5%, weight decay 0.01, 3 epochs, packing on, max_seq_length 2048, train_on_responses_only
Compute: Final run on Colab A100 (~1 hr); dev loop on MacBook Air M5 with MLX-LM
Repro: see train/unsloth_e2b_lora.py + pinned train/requirements-train.txt

Evaluation

Held-out gold sets (n≈50 each), entirely outside the training corpus.

Metric	n	Base E2B	E2B-ASHA	Δ
Protocol accuracy (Hindi)	25	24.0%	20.0%	-4.0 pp
Protocol accuracy (Marathi)	25	16.0%	12.0%	-4.0 pp
Function-call validity (tool schema in prompt)	15	100.0%	100.0%	0.0 pp
Refusal precision	20	85.0%	90.0%	+5.0 pp

Both models received the same Gemma-4-IT chat template + an inline ASHA-Saathi system prompt. Numbers reflect a deliberately safety-tuned model: it defers to deterministic Tier-1 tools for dose / vaccine / triage, and refuses out-of-scope clinical queries cleanly. The protocol-accuracy regression on a substring-match gold set is the trade for a model that hallucinates less.

See eval/results_v1.md for the latest run; on-device latency numbers land in eval/latency_v1.md once measured on a target Android device.

How to use

Transformers (server / desktop)

python

from transformers import AutoModelForCausalLM, AutoTokenizer

tok = AutoTokenizer.from_pretrained("ombhojane/gemma-4-e2b-asha-it")
mdl = AutoModelForCausalLM.from_pretrained("ombhojane/gemma-4-e2b-asha-it", torch_dtype="bfloat16", device_map="auto")

msgs = [{"role": "user", "content": "8 किलो के बच्चे को ORS कितना दें?"}]
ids = tok.apply_chat_template(msgs, return_tensors="pt", add_generation_prompt=True).to(mdl.device)
out = mdl.generate(ids, max_new_tokens=256, do_sample=False)
print(tok.decode(out[0][ids.shape[-1]:], skip_special_tokens=True))

Ollama (local CPU/GPU)

bash

ollama pull ombhojane/gemma-4-e2b-asha-it
ollama run ombhojane/gemma-4-e2b-asha-it "9 महीने के बच्चे का अगला टीका कौन सा है?"

On Android (the intended deployment)

Use the ASHA-Saathi APK which bundles the model, MediaPipe LLM Inference, the Tier-0 router, and Dart tool implementations.

Limitations & risks

Synthesis-derived training data. Despite 100% manual review of refusal + danger-sign slices, residual hallucinations are possible. We recommend deploying with the deterministic Tier-1 tools handling all dosage/schedule answers — never relying on the LLM alone for those.
Hindi/Marathi only. Generalization to other Indic languages is untested.
Not safety-certified. Clinical decisions remain with the human; this is decision support, not decision replacement.
Inherits Gemma 4 base limitations. Hallucination, prompt-injection susceptibility, etc.

License

Inherits Gemma's license terms (see Gemma usage policy).

Citation

bibtex

@model{gemma_4_e2b_asha_it_2026,
  author = {Bhojane, Om},
  title  = {gemma-4-e2b-asha-it: an offline ASHA co-pilot},
  year   = {2026},
  url    = {https://huggingface.co/ombhojane/gemma-4-e2b-asha-it}
}

Acknowledgements

Built on Gemma 4 by Google DeepMind. Trained with Unsloth. Deployed via MediaPipe LLM Inference / LiteRT. Submitted to the Gemma 4 Good Hackathon.