Private research dataset. Supporting data for MaternoVer, a multilingual (English / Hindi / Tamil) maternal-health RAG research project comparing WHO global guidance against India's National Health Mission (NHM) guidance, including source-grounded fact verification, cross-source conflict detection, and fine-tuning evaluation.

Author: Balaji Anbalagan (vijibalaji2003@gmail.com), independent research, supported by Dr. Selvi Ravindran and Dr. Bama S (Anna University, CEG IT Department).

What's in this dataset (6 configs)

Config	Records	What it is	Status
`who_corpus`	20	Hand-curated WHO Q&A pairs, forensically verified against primary WHO source PDFs (exact quote + page cited for every claim)	Final, audited
`nhm_corpus`	22	Hand-curated India NHM/MoHFW Q&A pairs, same forensic verification standard	Final, audited
`conflict_benchmark`	10	Labeled WHO-vs-NHM guidance pairs with a 5-way relationship label (`contradiction` / `compatible` / `protocol_divergence` / `different_scope` / `insufficient_evidence`)	Final, audited
`dev_benchmark`	237	79 base questions x 3 languages, 7 categories, full gold-answer schema -- a development/sandbox benchmark used to validate the question-design methodology	Dev set, not the final held-out test
`training_pilot`	118	Pilot fine-tuning examples across 8 behavior categories (citation-aware answering, refusal on unsupported questions, conflict-aware responses, safety escalation, etc.)	Pilot, not the full training set
`frozen300`	300	100 base questions x 3 languages (en/hi/ta), the frozen, held-out final test set	Final -- built, manually reviewed, translated, and independently (`agy`) spot-checked; see Provenance below

Why the "final" vs "pilot/dev" distinction matters

This is a research-integrity decision, not a formality: dev_benchmark was used to design and debug the question methodology (it caught and fixed real problems -- fake multi-hop questions, conflict-question phrasing that leaked the answer). frozen300 is the eventual held-out test set that methodology was designed to produce, and it is kept structurally separate so training data can never have seen it (checked against both dev_benchmark and training_pilot, 0 contamination flags on both). Do not use dev_benchmark questions as a substitute for a true held-out evaluation; it was explicitly built to be edited and was edited.

Provenance and verification methodology

Every who_corpus / nhm_corpus entry carries: source_document, source_organization, source_page, source_quote (the exact primary-source text), gold_source_id (a stable join key used across every other config in this dataset), provenance (primary_source or general_knowledge), and correction_status + correction_log documenting exactly what was verified, corrected, or found unsupported -- including several real misattributions caught and corrected during the audit (e.g. a widely-cited pregnancy weight-gain table commonly attributed to WHO that actually originates from the US Institute of Medicine).

An independent second-opinion validation pass (a different LLM, Google's Antigravity CLI, given web search access) was also run against the WHO corpus specifically to catch errors the original audit might have missed. It found one real, since-corrected error (a folic-acid preconception-timing claim originally miscategorized as CDC-only, actually confirmed as stated WHO guidance in a footnote the original audit missed). See correction_log entries and the project's own docs/PDF_GROUNDING_PROGRESS.md (not included in this dataset export) for the full audit trail.

frozen300's Hindi/Tamil translations went through the same kind of independent spot-check (agy, translation-fidelity mode) across all 200 language variants, plus a separate mechanical pattern checker. Every flag from both tools was manually verified against the live text before any fix was applied -- a meaningful share of the independent check's flags turned out to be false positives describing already-fixed bugs rather than the current text, so flags were treated as leads to verify, not as ground truth. 32 confirmed real translation errors were found and fixed this way (see docs/FROZEN300_BUILD_REPORT.md for the full list and reasoning behind each fix). 76 of the 100 questions were machine-translated (MyMemory) and QC'd per-batch; the remaining 24 were translated directly by an LLM (Claude) after a repeat translation-API quota block -- a disclosed methodology difference within the set, not a silent one.

Known limitations (stated plainly)

The 42-entry corpus (who_corpus + nhm_corpus) is author-compiled and source-grounded, not clinically validated by a licensed medical professional. Every claim traces to a real primary document, but this is not equivalent to a clinical guideline review.
dev_benchmark's Hindi/Tamil text was directly authored, not machine- translated; it has been back-translation-verified on a 20-question sample (mean similarity 0.692) but not the full set.
frozen300's translations have mixed provenance within the set (76 machine-translated, 24 LLM-translated -- see Provenance above); both groups passed the same verification bar, but this is a real methodology difference worth disclosing if the set is described in a paper.
training_pilot is a small pilot (118 examples), not the full training set a real fine-tuning run would use.

Intended use

Research and academic use in evaluating multilingual retrieval-augmented generation, cross-source conflict detection, and citation-grounded question answering in the maternal-health domain. Not intended for clinical deployment or as a substitute for professional medical advice.

License / usage note

See LICENSE.md. The underlying facts are drawn from WHO and Government of India (NHM/MoHFW) public guidance documents; this dataset's compiled question-answer structure, benchmarks, and annotations are the author's own work, shared here for research review, not for redistribution as an official WHO or NHM product.

Downloads last month: 3

Total file size:

1.08 MB