---
license: cc0-1.0
task_categories:
  - audio-classification
  - text-classification
language:
  - en
  - multilingual
tags:
  - medical
  - drug-names
  - rare-words
  - keyword-spotting
  - audio-text-embedding
size_categories:
  - 100K<n<1M
---

# Rare Medical Terms Dataset

A curated collection of **626,251 speakable medical/pharmaceutical/chemical terms** for training audio-text embedding models (e.g., CLAP-style keyword spotting).

## Sources

| Source | Terms | License |
|---|---|---|
| [DrugBank Vocabulary](https://go.drugbank.com/releases/latest#open-data) | ~33K | CC0 |
| [ChEBI](https://www.ebi.ac.uk/chebi/) | ~180K | Free/Open |
| [MeSH Descriptors](https://www.nlm.nih.gov/mesh/) | ~150K | Public Domain |
| [MeSH Supplementary Records](https://www.nlm.nih.gov/mesh/) | ~260K | Public Domain |

## Filtering

Raw terms were filtered to keep only "speakable" entries:
- No chemical formulas (brackets, equals signs, etc.)
- No numeric characters
- Length between 3-50 characters
- At least 90% alphabetic characters

## Format

JSONL with fields:
- `id`: integer index
- `term`: the medical/pharmaceutical term

## Use Case

Designed for fine-tuning audio-text contrastive models (Whisper encoder + text encoder with SigLIP loss) to detect rare medical terms in long audio via embedding similarity.