https://github.com/freedomintelligence/medpajama
MedPajama: A Large-scale Trustworthy Medical Corpus
https://github.com/freedomintelligence/medpajama
Last synced: 3 months ago
JSON representation
MedPajama: A Large-scale Trustworthy Medical Corpus
- Host: GitHub
- URL: https://github.com/freedomintelligence/medpajama
- Owner: FreedomIntelligence
- Created: 2025-06-05T18:25:13.000Z (12 months ago)
- Default Branch: main
- Last Pushed: 2025-06-06T04:13:41.000Z (12 months ago)
- Last Synced: 2025-10-12T05:04:45.255Z (8 months ago)
- Size: 137 KB
- Stars: 1
- Watchers: 1
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
README
## MedPajama: A Large-scale Trustworthy Medical Corpus
MedPajama is an ambitious initiative to build one of the largest, most reliable, and clinically credible medical corpora to support the development of next-generation medical AI systems.
Logo for MedPajama designed with ChatGPT-4o
## ⚡ Introduction
This project focuses on constructing a trustworthy medical dataset by:
1. Collecting and filtering massive-scale pretraining data from authoritative sources.
2. Annotating data with multi-level labels (e.g., medical domain, semantic type, source credibility).
3. Extracting the highest quality and most reliable data to build a RAG knowledge base.
4. Selecting high-confidence and challenging questions to construct datasets for SFT and RL.
Whether you're developing medical foundation models, disease-specific LLMs, or medical RAG systems, MedPajama provides a trusted data backbone to support your research and applications.
# 🎯 ToDo
- [ ] Collection of Pretraining Data and Domain-Specific Medical Filtering.
- [ ] Filter multi-dimensional labeled data and verify it with medical experts.
- [ ] Analyze how different levels of data trustworthiness impact medical LLM performance.
- [ ] Extract a subset of the highest-trust data to build a RAG knowledge base for retrieval.
- [ ] Select a subset of knowledge-rich or challenging data to construct SFT datasets.
- [ ] Select a subset of high-difficulty samples to construct RL training data.