SPerva/pillchecker-ner-benchmark

PillChecker NER Benchmark

Benchmark dataset for evaluating Named Entity Recognition (NER) models on pharmaceutical packaging text.

Dataset Description

500 synthesized pack-label texts generated from the MattBastar/Medicine_Details dataset, designed to simulate OCR output from photos of pill packaging.

Each case contains:

• id: Unique case identifier
• category: single_ingredient, dual_ingredient, or multi_ingredient
• ocr_text: Synthesized pharmaceutical label text (clean or with OCR noise)
• expected_names: Ground-truth list of active pharmaceutical ingredients
• source_composition: Original composition string from source dataset

Use Case

This dataset tests whether NER models can extract active pharmaceutical ingredients from short, formulaic packaging text — a domain significantly different from biomedical literature.

Benchmark Rules

This dataset is the canonical location for benchmark input cases only. Benchmark result history belongs in hf://buckets/SPerva/pillchecker-experiments/benchmark-results/, not in this dataset repository.

Current records contain id, category, ocr_text, expected_names, and source_composition. Before using this dataset for entity-linking or interaction claims, add expected_rxcuis, clean_text, expected_interactions, and known-safe pairs as described in the GitHub repo's AGENTS.md and eval/README.md.

Baseline Results

Historical OpenMed baseline results are stored in the PillChecker experiments bucket. Project-facing result claims should cite a bucket run manifest and Git commit.

Source

Part of the PillChecker project.