IndoNutriQA

An Indonesian-language clinical question answering (QA) dataset for the nutritional management of pediatric and adult inpatients. QA pairs are provided in SQuAD-style format, grounded in anonymized chunks extracted from real clinical nutrition case reports.

See DATASHEET.md for a full datasheet (Gebru et al., 2021).

• data/dataset/dataset.json — the QA dataset (questions, answers, and the supporting chunk IDs).
• data/chunks/ — anonymized context chunks, one JSON file per source case report. PII (names, dates, locations, registration numbers, hospital names) is replaced with <NAMA>, <TGL>, <LOKASI>, <NO_REG>, <RS>, <NO_PEGAWAI>, <NIM> tokens. Names of treating physicians, supervisors, report authors and presenters are also replaced.
• Additional folders under data/ contain evaluation artifacts (statistical robustness, semantic and RAGAS evaluations, categorical analyses).

Each entry in dataset.json is a SQuAD-style QA pair:

{
  "question": "...",
  "answer": "...",
  "context_chunks": ["chunk_id_1", "chunk_id_2"],
  "metadata": { ... }
}

Chunks referenced by context_chunks are stored in data/chunks/ and can be loaded by ID.

import json

with open("data/dataset/dataset.json") as f:
    qa_pairs = json.load(f)

• Data: CC BY-NC 4.0 — non-commercial use with attribution.
• Code: MIT.

Source case reports contained personally identifiable information (PII) that has been removed via an NER-based anonymization pipeline. Users must not attempt to re-identify individuals. See DATASHEET.md for details.

The saved predictions and results under data/ are enough to recompute every table and figure of the paper without calling any LLM API again. The scripts read the dataset from data/dataset/dataset.json.

python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt

To call LLM APIs again, create a .env file at the project root with OPENAI_API_KEY=... and ANTHROPIC_API_KEY=....

• The original report PDFs, the answer spreadsheets (QnA.xlsx) and the intermediate extraction steps. They contain patient, hospital and staff identifiers. Pipeline steps 1 to 4c are included only to document how the data was prepared and cannot be re-run from this repository.
• Fine-tuned model checkpoints. The saved predictions in data/13-statistical_robustness/predictions/ are enough to recompute all scores.

src/pipelines/        numbered pipeline steps
src/utils/            PDF processing, anonymisation, validation, Hugging Face formatting
config/               shared settings
data/dataset/         the QA dataset
data/chunks/          anonymised context chunks
data/6-... to 20-...  saved predictions and results, one folder per step
docs/rd_annotation/   rubric and labelling set-up for the expert-rater study
notebooks/            demographic and anthropometric figures

Run every script from the project root, for example python src/pipelines/12-llm_evaluation.py --help.

The comparison tables with other benchmarks were compiled by hand from the original papers and have no script.

If you use IndoNutriQA, please cite the accompanying paper (citation details to follow).