Saminx22/medical_data_for_slm

🏥 Medical SLM Pretraining Dataset Card

This dataset is a high-quality, cleaned collection of medical text designed for pretraining small language models (SLMs). It aggregates data from three primary authoritative sources, focusing on general medicine and clinical guidelines.

📊 Dataset Summary

• Total Documents: ~44,400
• Estimated Tokens: ~44.7 Million
• Primary Language: English
• Configurations:
  • documents: Raw cleaned text records.
  • chunks: Tokenized and packed 1024-token sequences for training.

🛠 Processing Pipeline

1. Cleaning: Removed boilerplate (copyrights, funding, URLs, DOIs) and normalized whitespace.
2. Quality Filtering: Removed documents that were too short (<300 chars), had low alphabetic ratios, or high digit/special character ratios.
3. Deduplication: Applied exact MD5 hashing to remove identical documents.
4. Packing: Greedy packing of documents into 1024-token chunks separated by <|endoftext|>Requested.

📚 Original Datasets & Credits

We gratefully acknowledge the creators of the original datasets used to compile this corpus:

1. PubMed Abstracts

• Source: NCBI PubMed
• Credit: National Library of Medicine (NLM).
• Usage: Scientific abstracts summarizing biomedical research.

2. PMC Open Access (PMC OA)

• Source: PMC Open Access Subset
• Credit: National Institutes of Health (NIH) / PubMed Central.
• Usage: Full-text articles providing deep clinical context.

3. Clinical Guidelines

• Source: EPFL-LLM Guidelines
• Credit: EPFL-LLM Team. Aggregated from WHO, CDC, NICE, and other health organizations.
• Usage: Authoritative medical standards and clinical practice protocols.

⚖️ License

This aggregate dataset is provided for research purposes. Users must adhere to the individual licenses of the source datasets (Creative Commons, NLM, etc.).