---
license: cc-by-4.0
language:
  - hi
  - mr
size_categories:
  - 1K<n<10K
task_categories:
  - question-answering
  - text-generation
tags:
  - medical
  - india
  - asha
  - maternal-health
  - child-health
  - instruction-tuning
  - hindi
  - marathi
pretty_name: ASHA-Saathi Instructions (Hindi + Marathi) v1
---

# ASHA-Saathi Instructions (Hindi + Marathi) v1

A reusable Indic-medical instruction-tuning dataset for fine-tuning small language models to assist **ASHA workers** — India's ~1 million government-employed frontline community-health workers — in **Hindi** and **Marathi**.

> Built as the training corpus for [`ombhojane/gemma-4-e2b-asha-it`](https://huggingface.co/ombhojane/gemma-4-e2b-asha-it), submitted to the [Gemma 4 Good Hackathon](https://www.kaggle.com/competitions/gemma-4-good-hackathon). Released independently so other researchers can fine-tune **any** small Indic-language LM on the same task.

## Quick load

```python
from datasets import load_dataset

ds = load_dataset("ombhojane/asha-instructions-hi-mr-v1")
print(ds)
# DatasetDict({
#   train: Dataset({features: [...], num_rows: ~5800})
#   validation: Dataset({...})
#   test: Dataset({...})
# })
```

## Why this exists

ASHA workers serve ~250 households each, often offline, with low-end Android phones. Their daily decisions are protocol-grounded (vaccine schedules, ORS dosing, danger-sign triage, ANC/PNC checklists), but no open instruction-tuning corpus exists for this exact slice. Generic Indic-instruction datasets (Aya, indic-align) cover language well but not medical-protocol fidelity. Generic medical datasets (MedMCQA, PubMedQA) cover medicine well but not Hindi/Marathi or ASHA scope.

This dataset bridges both.

## Composition

| Slice | % | What it teaches |
|---|---|---|
| `protocol_qa` | 60% | Q&A grounded in NHM ASHA Modules 1–7, IMNCI, ICMR maternal-health protocols |
| `function_call` | 25% | Multi-turn tool-use for `dosage_calculator`, `vaccine_schedule`, `danger_sign_check`, `nearest_phc_referral` |
| `refusal` | 10% | Out-of-scope refusal — ASHA should NOT diagnose cancer, prescribe antibiotics, etc. — refer up |
| `danger_sign` | 5% | High-stakes triage — recognize IMNCI danger signs and escalate immediately |

Hindi and Marathi are split evenly within each slice. Train/validation/test split is 90/5/5, **stratified by `(slice, lang)`** with seed 42.

## Schema

Each row is one example:

```json
{
  "instruction": "बच्चे का वजन 6 किलो है, ORS कितना देना चाहिए?",
  "input": "",
  "output": "6 किलो वज़न के बच्चे को हर दस्त के बाद 75 ml ORS देना चाहिए...",
  "turns": [],
  "lang": "hi",
  "slice": "protocol_qa",
  "source": "ASHA_Module_3_Hi#chunk12",
  "reviewed": true,
  "reviewer": "om",
  "notes": "",
  "seed": 42,
  "teacher_model": "gemini-2.5-pro",
  "teacher_temperature": 0.7
}
```

For `function_call` examples, `turns[]` carries the full multi-turn trajectory (user → assistant tool_call → tool result → assistant reply). `instruction` and `output` always carry the user's first message and the assistant's final reply for compatibility with simple SFT loaders.

## How it was built

1. **Source extraction** — NHM ASHA Modules 1–7 (Hindi), WHO IMNCI, ICMR maternal-health guidelines, vikaspedia.in MCH module (Marathi). All public-license / GoI-published.
2. **Synthesis** — A teacher LLM (Gemini 2.5 Flash, Gemma 4 31B for the smaller refusal/danger-sign slices where rate limits permit) authored Q&A from passages with explicit grounding constraints. We reject any answer whose token-overlap with its source passage falls below 30%. Cross-family distillation (Gemini teacher → Gemma 4 E2B student) is a common, well-documented pattern; the choice was driven by the teacher's API throughput, not capability.
3. **Manual review** — 100% of `refusal` and `danger_sign` slices are human-reviewed (`reviewed=true`, `reviewer` field set). Other slices are sampled-reviewed at 10%.
4. **Validation** — Function-call examples are validated against tool JSON schemas; invalid calls are dropped.
5. **Deduplication** — by `(lang, slice, instruction.lower())` SHA-256 fingerprint.

The full pipeline lives in [`data/build_instruction_set.py`](https://github.com/ombhojane/asha-saathi/blob/main/data/build_instruction_set.py) — fully reproducible from the same seed.

## License

CC-BY-4.0. You may use this dataset commercially with attribution.

Source-licence breakdown:
- NHM ASHA Modules — Government of India (GoI) public-domain
- WHO IMNCI / IMCI — WHO public-use
- vikaspedia.in — CDAC, attribution required
- Synthetic content — author-released under CC-BY-4.0

## Limitations

- **Not a clinical reference.** Models trained on this dataset are training aids for ASHAs, not replacements for clinical judgment.
- Synthesized via teacher LLM — minor stylistic biases may persist despite review.
- Currently Hindi + Marathi only. Tamil/Bengali/Telugu planned for v2.
- The `function_call` `tool_result` field is plausible-but-synthetic; real deployments must wire actual deterministic tool outputs.

## Citation

```bibtex
@dataset{asha_saathi_v1_2026,
  author = {Bhojane, Om},
  title  = {ASHA-Saathi Instructions (Hindi + Marathi) v1},
  year   = {2026},
  url    = {https://huggingface.co/datasets/ombhojane/asha-instructions-hi-mr-v1}
}
```

## Maintainers

- [Om Bhojane](https://huggingface.co/ombhojane) — corpus design, review

Issues / contributions welcome on [github.com/ombhojane/asha-saathi](https://github.com/ombhojane/asha-saathi).

## Changelog

- **v1.0.0** — Initial release for the Gemma 4 Good Hackathon.