---
license: apache-2.0
library_name: transformers
pipeline_tag: image-text-to-text
datasets:
- rsoohyun/SpatialBlock-15k
base_model:
- Qwen/Qwen3-VL-4B-Instruct
---

# SpatialBlock-4B-direct

This repository contains the **SpatialBlock-4B-direct** checkpoint from the paper [SpatialBlock: Enhancing Spatial Intelligence in LVLMs via Synthetic Block-Stacking Problem](https://huggingface.co/papers/2609.07064).

It is a fine-tuned version of [Qwen3-VL-4B-Instruct](https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct) on the synthetic [SpatialBlock-15k](https://huggingface.co/datasets/rsoohyun/SpatialBlock-15k) dataset. The model directly predicts answers to spatial reasoning tasks such as 3D-to-2D projection, viewpoint transformation, and structural combination.

For training details, evaluation results, and the companion “reason” model, please refer to the GitHub repository: [https://github.com/rsoohyun/SpatialBlock](https://github.com/rsoohyun/SpatialBlock).