Pharmaceutical regulatory intelligence aggregated from 200 countries and 202 regulatory authorities worldwide. Built for retrieval-augmented generation over global medicines regulation.

Dataset summary

Metric	Value
Countries covered	200
Regulatory authorities	202
Documents	12,533
RAG chunks	30,768
Total size	273.2 MB
Created	2026-04-28
Schema version	1.0

Data sources

Source	Documents
Direct web scraping (national authority sites)	10,360
API integrations (FDA, OpenFDA, etc.)	8,966
Secondary sources (gap-fill from WHO, regional bodies, peer references)	142

Layout

.
├── README.md                                   (this file)
├── all_countries/
│   ├── documents_all.jsonl                     (all 12,533 documents)
│   └── chunks_all.jsonl                        (all 30,768 chunks)
├── by_country/
│   ├── Afghanistan/
│   │   ├── documents.jsonl
│   │   └── chunks.jsonl
│   ├── Albania/
│   ├── ... (200 country folders)
│   └── Zimbabwe/
└── hf_export_summary.json

Document schema

Each line in documents.jsonl is a single JSON object:

{
  "document_id": "uuid",
  "source": {
    "country": "AZ",
    "country_name": "Azerbaijan",
    "authority": "ASMP",
    "authority_full_name": "Agency for Standardization Metrology and Patents",
    "tier": 3,
    "url": "https://...",
    "scrape_timestamp": "2026-04-28T..."
  },
  "content": {
    "title": "...",
    "content_type": "GUIDELINE | LEGISLATION | PUBLICATION | ...",
    "category": "HUMAN_MEDICINES | MEDICAL_DEVICES | ...",
    "subcategory": "",
    "language": "en",
    "original_language": "en",
    "is_translated": false
  },
  "freshness_metadata": {
    "document_date": "...",
    "effective_date": "...",
    "scrape_date": "2026-04-28",
    "status": "ACTIVE",
    "freshness_score": "CURRENT"
  },
  "text": "...",
  "sections": [{ "heading": "...", "level": 1, "content": "..." }],
  "tags": ["legislation", "human-medicines", "pharmacovigilance"],
  "word_count": 297,
  "chunk_ready": true
}

Chunk schema

Each line in chunks.jsonl is one chunk ready for vector-store ingestion:

{
  "chunk_id": "uuid",
  "document_id": "parent uuid",
  "chunk_index": 0,
  "total_chunks": 5,
  "context_prefix": "[Country: ... | Authority: ...]",
  "text": "...",
  "token_count": 512,
  "metadata": {
    "country": "AZ",
    "authority": "ASMP",
    "tier": 3,
    "content_type": "LEGISLATION",
    "category": "HUMAN_MEDICINES",
    "language": "en",
    "status": "ACTIVE"
  }
}

Authority tiers

Tier 1: WHO Listed Authorities (WLAs) and EU Medicines Regulatory Network — mature, high-capacity regulators (FDA, EMA, MHRA, PMDA, MFDS, HSA, Swissmedic, etc.).
Tier 2: ML3/ML4 / transitional WLAs — well-functioning regulators in major emerging economies (ANVISA, CDSCO, NMPA, ANMAT, etc.).
Tier 3: National regulators across the rest of the world.

Coverage methodology

The dataset combines four content acquisition strategies:

Direct scraping of national authority websites (Tier 1, 2, and 3 with available URLs).
API integration with FDA OpenFDA, ICH, WHO, and regional harmonisation bodies.
Connection-based gap filling for countries without scrapable national portals — pulls from WHO Global Benchmarking Tool data, regional regulatory cooperation memberships (EAC, ASEAN, GCC, CIS, Council of Europe, Arab League, etc.), and recognised reference-authority relationships (FDA, EMA, WHO-PQ, TGA).
Recovery pass for countries that were scraped but produced no accepted documents — uses the same connection-based approach.

Pipeline

Documents pass through a 5-stage pipeline before inclusion:

SCRAPE → CLEAN → NORMALIZE → ENRICH → VALIDATE → EXPORT

Clean: HTML→markdown, noise removal, mojibake fixes.
Normalize: ISO-codification of country names, date standardisation, language detection.
Enrich: Freshness scoring, category classification, related-document linking, tag generation.
Validate: Length, scope (human medicines), freshness, deduplication (SHA-256 + SimHash), language, link checks.
Export: Semantic chunking with cl100k_base tokenizer at 512 tokens / 64 overlap.

Pipeline config

Parameter	Value
Chunk size (tokens)	512
Chunk overlap (tokens)	64
Tokenizer	cl100k_base
Classifier threshold	0.6
Dedup method	simhash
Dedup threshold	0.95

Loading the dataset

from datasets import load_dataset

# Documents only
docs = load_dataset("RuthvikBandari/Zebrafish_Countries_data", split="documents")

# RAG chunks
chunks = load_dataset("RuthvikBandari/Zebrafish_Countries_data", split="chunks")

# Or stream a single country
import pandas as pd
azerbaijan = pd.read_json(
    "hf://datasets/RuthvikBandari/Zebrafish_Countries_data/by_country/Azerbaijan/documents.jsonl",
    lines=True,
)

Limitations

Coverage breadth is uneven: Tier 1 authorities (FDA, EMA, etc.) dominate document count; many small/island/least-developed nations rely on connection-based gap-fill content rather than direct scrapes.
Gap-fill records are clearly marked: their URLs use the urn:zebrafish:peer-reference:<ISO>, urn:zebrafish:who-profile:<ISO>, urn:zebrafish:ich-adoption:<ISO>, or urn:zebrafish:gap-fill:<ISO> URN namespace. Real scrapes have https:// URLs.
Content currency: documents reflect the state of authority websites at the scrape timestamps recorded per document. Regulatory information evolves; verify against the authority website before relying on a specific provision.
Languages: the dataset is primarily in English. Some non-English source pages were scraped in their original language and tagged accordingly.

License

Document content is sourced from public regulatory authority websites and international organisation publications. Each document retains its source URL for attribution; users are responsible for compliance with the source authority's terms of use.

Pipeline code, gap-fill content, and dataset assembly are provided under the project's existing license terms.

Citation

@dataset{zebrafish_rag_2026,
  title  = {Zebrafish Global Drug Regulatory RAG Dataset},
  author = {Ruthvik Bandari},
  year   = {2026},
  url    = {https://huggingface.co/datasets/RuthvikBandari/Zebrafish_Countries_data}
}

Downloads last month: 9

Total file size:

573 MB