ekacare/Eka_NFI_MCQA

Eka NFI MCQA

A multiple-choice question answering dataset for evaluating knowledge of drug pharmacology as defined in the National Formulary of India (NFI) 2011 — covering indications, contraindications, dosing, precautions, pregnancy safety, and drug schedules.

Dataset Description

This dataset contains 925 questions derived from the National Formulary of India (NFI), 2011, a government-published reference for drug information in the Indian healthcare context. Questions span 6 pharmacological categories and test both factual recall and clinical reasoning.

Use Cases

• Evaluate medical LLMs on drug knowledge grounded in the Indian national formulary
• Benchmark pharmacology understanding across indications, contraindications, dosing, and safety
• Test drug schedule and pregnancy category classification per Indian regulatory standards

Dataset Structure

Column	Type	Description
question	string	The MCQ question stem
options	string	JSON-encoded dict of answer options (keys: A–E)
answer	string	Correct answer letter (A–E)
explanation	string	Explanation of the correct answer
question_type	string	Fine-grained question type (e.g., dose_mcq, contraindications_multi_select)
category	string	Pharmacological category (see below)
difficulty	string	Difficulty level (medium or hard)
generation_timestamp	string	ISO timestamp of question generation

Categories

Category	Questions	Description
contraindications	166	Conditions or patient profiles where a drug must not be used
precautions	166	Situations requiring caution or dose modification
pregnancy_category	166	Drug safety classification during pregnancy
dose	157	Dosing regimens, routes, and frequencies
schedule	148	Drug schedule classification under Indian law
indications	122	Approved therapeutic uses of the drug

Question Types

Each category includes both single-correct (_mcq) and multi-select (_multi_select) formats:

Question Type	Count
contraindications_mcq	164
precautions_mcq	124
dose_mcq	156
schedule_mcq	148
pregnancy_category_mcq	148
indications_mcq	121
pregnancy_safety_mcq	18
precautions_multi_select	42
contraindications_multi_select	2
dose_multi_select	1
indications_multi_select	1

Difficulty Distribution

Difficulty	Count
medium	911
hard	14

How It Was Made

1. Source extraction — The NFI 2011 PDF was parsed to extract structured drug annotations (indications, contraindications, dosing, precautions, pregnancy categories, schedules).
2. Human review — Extracted annotations were reviewed and validated by humans for accuracy and completeness.
3. LLM-generated MCQA — Questions and distractor options were generated by an LLM using the validated annotations as ground truth.
4. Human review — Generated questions were reviewed again to ensure clinical correctness and remove low-quality items.

Usage

With KARMA Evaluation Framework

karma eval \
  --model "Qwen/Qwen3-0.6B" \
  --datasets "ekacare/Eka_NFI_MCQA" \
  --format table

Direct Loading

from datasets import load_dataset
import json

ds = load_dataset("ekacare/Eka_NFI_MCQA", split="test")

example = ds[0]
print(example["question"])
print(json.loads(example["options"]))  # {"A": "...", "B": "...", ...}
print("Answer:", example["answer"])
print("Explanation:", example["explanation"])

Evaluation

For single-correct questions (_mcq types), use exact match on the predicted answer letter (A–E).

For multi-select questions (_multi_select types), the correct answer may be a combination letter like "A, C" — parse accordingly.

def score(prediction, ground_truth):
    return int(prediction.strip().upper() == ground_truth.strip().upper())

Citation

If you use this dataset, please cite:

@dataset{eka_nfi_mcqa_2025,
  title={Eka NFI MCQA},
  author={Eka.Care},
  year={2025},
  publisher={HuggingFace},
  url={https://huggingface.co/datasets/ekacare/Eka_NFI_MCQA}
}

License

This dataset is released under CC-BY-NC-4.0. The source material is the National Formulary of India, 2011, published by the Ministry of Health and Family Welfare, Government of India.