Skip to content

Dataset · microsoft

Microsoft medical instructions for SFT and DPO

microsoft/mediflow

A synthetic dataset of 2.5 million clinical instructions in English for SFT and DPO alignment.

What's inside

Contains 2.5 million rows (about 700 thousand unique instructions) and tabular data in English for clinical NLP, covering 14 task types and 98 medical document types. The mediflow split contains data for SFT, while mediflow_dpo contains about 135 thousand top instructions with rejected responses from GPT-4o for DPO.

What you can do with it

  • SFT fine-tuning of medical language models and LoRA adapters
  • DPO alignment of LLMs on clinical tasks
  • Automating the processing and analysis of medical documents
Tasks
Text generation
Languages
English
Size
1M–10M rows
Downloads this month
333
Likes
54

Models trained on it

How to load it

from datasets import load_dataset

ds = load_dataset("microsoft/mediflow")