---
language:
- en
task_categories:
- visual-question-answering
- image-to-text
tags:
- clevr
- change-detection
- multi-sentence
license: mit
---

# CLEVR-Change Sentence-Balanced Dataset

This dataset is a processed subset of **Lancelot53/clevr-change** with an explicit
`count` feature.

## Construction

- Each sample contains:
  - `sentences: List[str]`
  - `sentence_count: int`
  - paired images (`image_1`, `image_2`)
- We construct a balanced subset with **50 samples for each sentence count from 3 to 9**.

## Features

- `sentences`: list of sentences in the query
- `sentence_count`: number of sentences (verified by sanity check)
- other fields inherited from CLEVR-Change

## Intended Use

- Analyzing reasoning difficulty in multi-sentence visual change understanding
- Curriculum learning and controlled evaluation for VLMs

## Source Dataset

- Original dataset: https://huggingface.co/datasets/Lancelot53/clevr-changen-*
---