Skip to content

Dataset · RuoliuYang

Visual Latent Reasoning Subsets (ULVR)

RuoliuYang/ulvr_subset

Curated subsets of 100k to 1M examples featuring images, dialogues, and precomputed teacher latents.

What's inside

Curated nested subsets of the first training stage of Unified Visual Latent Reasoning (ULVR) with a volume of 100K to 1M records. Each folder contains source data (images, question-answer pairs, and message-format dialogues) and their corresponding precomputed latent representations from the teacher model, linked by sample_id.

What you can do with it

  • Training and distilling multimodal models for visual question answering tasks
  • Experimenting with latent reasoning in image-to-text models
Tasks
Visual question answering, Image → text
Size
100K–1M rows
Downloads this month
32.7K
Likes
0

How to load it

from datasets import load_dataset

ds = load_dataset("RuoliuYang/ulvr_subset")