InnovatorLab/EMVista

---
license: mit
task_categories:
- visual-question-answering
language:
- en
tags:
- multimodal
pretty_name: EMVista
size_categories:
- 1K<n<10K
configs:
- config_name: default
data_files:
- split: test
path: data/test.parquet
---

# EMVista Dataset
<center><h1>EMVista</h1></center>

<p align="center">
<img src="./assets/pipeline.png" alt="EMVista" style="display: block; margin: auto; max-width: 70%;">
</p>

<p align="center">
| <a href="https://huggingface.co/datasets/EMVista/EMVista"><b>HuggingFace</b></a> |
<a href="https://huggingface.co/papers/2601.19325"><b>Paper</b></a> |
<a href="https://github.com/InnovatorLM/Innovator-VL"><b>Code</b></a> |
</p>

---

## 🔥 Latest News

- **[2026/01]** EMVista v1.0 is officially released.

<!-- <details>
<summary>Unfold to see more details.</summary>
<br>

- EMVista supports **English** prompts.

</details> -->

<!-- ---

## Motivation: TODO

<details>
<summary>Unfold to see more details.</summary>
<br>

Recent advances in Multimodal Large Language Models (MLLMs) have demonstrated impressive performance on generic vision-language benchmarks. However, most existing benchmarks primarily assess **coarse-grained perception** or **commonsense visual understanding**, falling short in evaluating models’ abilities to reason over **complex, expert-level visual information**.

In realistic applications—such as scientific analysis, technical inspection, diagram interpretation, and abstract visual reasoning—models must go beyond recognizing objects or captions. They need to **extract structured visual cues**, **understand implicit visual attributes**, and **perform multi-step reasoning across multiple visual sources**.

To address this gap, we introduce **EMVista**, a benchmark designed to systematically evaluate multimodal models’ **visual understanding and reasoning capabilities** through carefully curated expert-level visual tasks.

</details>

--- -->
## Overview
**EMVista** is a benchmark for evaluating **instance-level microstructural understanding** in electron microscopy (EM) images across **three core capability
dimensions**:

1. **Microstructural Perception**
Evaluates the ability to detect, delineate, and separate individual
microstructural instances in complex EM scenes.
2. **Microstructural Attribute Understanding**
Measures the capacity to interpret key microstructural attributes, including
morphology, density, spatial distribution, layering, and scale variation.
3. **Robustness in Dense Scenes**
Assesses model stability and accuracy under extreme instance crowding,
overlap, and multi-scale complexity.

EMVista contains **expert-annotated EM images** with instance-level labels and
structured attribute descriptions, designed to reflect **realistic challenges**
in materials microstructure analysis.

---
## Dataset Characteristics

- **Task Format**: Visual Question Answering (VQA)
- **Modalities**: Image + Text
- **Languages**: English
- **Annotation**: Expert-verified
---

### Download EMVista Dataset

You can download the EMVista dataset using the HuggingFace `datasets` library
(make sure you have installed
[HuggingFace Datasets](https://huggingface.co/docs/datasets/quickstart)):

```python
from datasets import load_dataset

dataset = load_dataset("InnovatorLab/EMVista")
```

## Evaluations

We use [lmms-eval](https://github.com/EvolvingLMMs-Lab/lmms-eval) for evaluations. Please see [here](./evaluation/README.md) for detail files.

## License

EMVista is released under the MIT License. See [LICENSE](./LICENSE) for more details.

## Citation

```bibtex
@article{wen2026innovator,
title={Innovator-VL: A Multimodal Large Language Model for Scientific Discovery},
author={Wen, Zichen and Yang, Boxue and Chen, Shuang and Zhang, Yaojie and Han, Yuhang and Ke, Junlong and Wang, Cong and others},
journal={arXiv preprint arXiv:2601.19325},
year={2026}
}
```