Skip to content

Dataset · ZTE-AIM

Multimodal Data for Curr-ReFT Model RL Fine-Tuning

ZTE-AIM/Curr-ReFT-data

Multiple-choice visual question answering data for two-stage VLM post-training.

What's inside

The set contains 1,000 to 10,000 image examples for visual question answering (VQA) with multiple choice. The data is intended for two consecutive training stages of the Curr-ReFT model lineup (3B and 7B): Curriculum RL and sample selection based on rejected answers. The dataset was published by the ZTE-AIM team along with code and a paper.

What you can do with it

  • Post-training vision-language models (VLMs) via Curriculum RL
  • Training LoRA adapters to improve visual reasoning
  • Reproducing and adapting the Curr-ReFT model fine-tuning pipeline
Tasks
Question answering, Multiple choice
Size
1K–10K rows
Downloads this month
43.5K
Likes
4

Models trained on it

How to load it

from datasets import load_dataset

ds = load_dataset("ZTE-AIM/Curr-ReFT-data")