Skip to content

MedTrinity-25M

A large-scale multimodal medical dataset with multi-granularity annotations

Visit official site

What MedTrinity-25M is

The dataset is described as a large-scale multimodal collection with multi-granularity medical annotations, developed by multiple authors to advance research on medical image and text processing. Its construction covers data extraction and multi-granularity text description generation, and it supports medical image analysis tasks such as visual question answering and pathology image analysis.

What you can do with it

  • Extract metadata, region locations and medical knowledge from collected data
  • Prompt a large language model to generate fine-grained annotations
  • Run training and evaluation scripts for pretraining and fine-tuning
  • Fine-tune pretrained models such as LLaVA-Med++
  • Follow the quick start guide for installation and usage
  • Read the arXiv paper for research background and methods

Application scenarios

  • Train a deep learning model that identifies lesions in medical images
  • Develop a system that generates medical image reports automatically
  • Use the dataset as teaching material for medical image analysis

Getting started

  • Visit the GitHub page and clone the dataset locally
  • Install the required packages and dependencies from the quick start guide
  • Download and install the LLaVA-Meta-Llama-3-8B-Instruct-FT-S2 base model
  • Run pretraining and fine-tuning with the provided scripts

Pros & cons

✓ What we like

  • Includes a quick start guide with installation steps
  • The model library supports fine-tuning on medical image tasks

! What to watch out for

  • Fine-grained annotations are generated by prompting a large language model
  • Setup requires cloning the dataset and installing a base model

FAQ

What tasks are named?

Visual question answering and pathology image analysis are listed among supported medical image analysis tasks.

What comes with the dataset?

The material lists training and evaluation scripts, pretrained models such as LLaVA-Med++, a quick start guide and a paper on arXiv.

Last reviewed: 2026-11-15

More AI medical tools

View all →

How we review