Tushar9802/sakhi-asha-home-visit-conversations

Sakhi — ASHA Home-Visit Conversations (Hindi/Hinglish → Structured Forms)

Synthetic Hindi/Hinglish conversations between an Indian ASHA (Accredited Social Health Activist) and a patient during a maternal- and child-health home visit, each paired with a structured JSON target. Built for the Sakhi project — an offline voice-to-form tool for ASHA workers (github.com/Tushar-9802/Sakhi).

The dataset supports two supervised tasks over the same conversations:

• —`form_extraction` — extract a visit register (vitals, assessments, counselling, symptoms) into a visit-type-specific JSON schema.
• —`danger_signs` — detect NHM-defined clinical danger signs and produce a referral decision, with a verbatim quote as evidence for each flagged sign.

Why this exists

ASHA workers fill RCH (Reproductive & Child Health) registers by hand after each home visit. The conversation happens in spoken Hindi mixed with English clinical terms ("BP", "IFA", "TT") and numbers spoken as words. There was no open dataset pairing such conversations with the structured register fields they map to. This is that pairing.

Format

JSONL, one visit per line, in chat-message format:

{
  "messages": [
    {"role": "system",    "content": "<task instructions + strict null/evidence rules>"},
    {"role": "user",      "content": "<ASHA/patient conversation + output JSON schema>"},
    {"role": "assistant", "content": "<target JSON>"}
  ],
  "metadata": {
    "task": "form_extraction" | "danger_signs",
    "visit_type": "anc_visit" | "pnc_visit" | "delivery" | "child_health",
    "schema": "<schema name, form_extraction rows only>",
    "has_danger_signs": true | false,
    "source_id": "<stable id>"
  }
}

The system and user messages embed the exact instructions and JSON schema used at training time, so each row is self-contained and usable directly for SFT.

Composition

Visit-type coverage (train / validation):

Of the danger_signs rows, roughly 42% contain at least one danger sign and the rest are routine visits with an empty danger_signs array — deliberately weighted toward the no-danger case, since most real home visits have none.

Design notes

The targets are written to train null-strict, evidence-grounded extraction:

• —Fields not stated in the conversation are null — no guessing from context.
• —Names, dates, phone numbers and addresses are never invented; an unnamed patient ("दीदी", "बहन") maps to name: null.
• —Every flagged danger sign carries a utterance_evidence field that is a verbatim copy from the conversation, not a paraphrase.
• —Numbers spoken as Hindi words ("एक सौ दस बटा सत्तर") map to digits ("110/70").

Schemas are grounded in India's NHM RCH register and HBNC visit-day structure.

Provenance & privacy

All conversations are synthetic — generated for this project, not transcripts of real visits. No real patient, ASHA worker, or facility is represented. Patients are referred to only by kinship terms, so the data contains no personal identifiers.

Limitations

• —Synthetic, not field-recorded. Phrasing is cleaner and more on-topic than real spoken Hindi; it does not capture ASR noise, disfluency, or regional dialect variation.
• —Hindi-centric. Code-switching is Hindi + English clinical terms only; other Indian languages are out of scope.
• —Not a clinical reference. Danger-sign and referral labels follow NHM guidance for a data-extraction task. This dataset must not be used to make clinical decisions.
• —A small fraction of the original raw generations had schema leakage; 454 such issues were corrected before this release.

Intended use

Supervised fine-tuning and evaluation of small/medium LLMs for structured clinical-form extraction and rule-grounded danger-sign detection from code-switched Hindi text.

Citation

@misc{sakhi_asha_conversations_2026,
  title  = {Sakhi: ASHA Home-Visit Conversations (Hindi/Hinglish to Structured Forms)},
  author = {Tushar},
  year   = {2026},
  url    = {https://github.com/Tushar-9802/Sakhi}
}

License

CC-BY-4.0.