README.md

metadata

license: apache-2.0
task_categories:
  - image-segmentation
pipeline_tag: any-to-any
library_name: bagel-mot
tags:
  - sgt
  - semantic-generative-tuning
  - unified-multimodal
  - visual-understanding
  - visual-generation

SAM-SGT: Segmentation Training Data for Semantic Generative Tuning

This repository contains the SAM-SGT dataset, a collection of approximately 190k segmentation samples derived from SAM. It was introduced in the paper Semantic Generative Tuning for Unified Multimodal Models.

Project Page | Paper | Code

Dataset Description

SGT (Semantic Generative Tuning) is a training paradigm that couples visual understanding and generation in Unified Multimodal Models (UMMs) by using image segmentation as a generative proxy. This dataset provides high-level semantic supervision used to align multimodal representation spaces.

• Content: ~190,000 segmentation samples sourced from the Segment Anything (SAM) dataset.
• Format: Tar-sharded.
• Role: Serves as a high-level semantic proxy task to enhance vision-centric perception and generative layout fidelity.

Training Data Distribution

Data Source	Samples
SGT — Segmentation (SAM)	190k
General VQA	180k
Doc / Chart / Screen	103k
Math / Reasoning	101k
Language	72k
General OCR	45k
Total	~691k

Usage

You can download the dataset using the provided script from the official repository:

# download sam subset || Chinese users can use --use-mirror
python download_sam.py --target-dir ./data/SAM-SGT --use-mirror

Citation

If you find this work useful, please cite:

@article{yu2026sgt,
  title     = {Semantic Generative Tuning for Unified Multimodal Models},
  author    = {Yu, Songsong and Chen, Yuxin and Shan, Ying and Li, Yanwei},
  journal   = {arXiv preprint arXiv:2605.18714},
  year      = {2026},
}