README.md

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

# MolParse Bench

<center><h1>MolParse</h1></center>

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

<p align="center">
  | <a href="https://huggingface.co/datasets/InnovatorLab/MolParse"><b>HuggingFace</b></a> |
</p>

---

## 🔥 Latest News

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

---

## Overview

**MolParse** is a large-scale multimodal dataset for **optical chemical structure parsing**, designed to evaluate and train models that convert **molecular structure images** into **structured chemical representations**.

The dataset focuses on realistic chemical diagrams commonly found in scientific literature and patents, emphasizing robustness to visual noise, diverse drawing styles, and complex molecular layouts. MolParse supports tasks that require precise visual perception and structured chemical understanding, rather than simple text recognition.

---

## Benchmark Scope

MolParse evaluates models across the following core capability dimensions:

### 1. Molecular Structure Perception

Assess the ability to accurately recognize:
- Atoms and bonds
- Ring systems and fused structures
- Substituents and functional groups
- Variable attachment points and abstract structures

### 2. Structured Chemical Representation

Evaluate the capacity to translate molecular images into:
- Linearized chemical strings
- Structured symbolic representations
- Machine-readable formats suitable for downstream reasoning

### 3. Robustness in Real-World Documents

Test model stability under:
- Noisy or low-quality scans
- Diverse drawing conventions
- Crowded layouts and overlapping annotations
- Variations in resolution and aspect ratio

---

## Dataset Characteristics

- **Task Format**: Image-to-Structure Parsing
- **Modalities**: Image + Text
- **Domain**: Chemistry
- **Languages**: English
- **Annotation**: Expert-verified
- **Data Scale**: Large-scale (millions of image–structure pairs)

---

## Task Types

Each MolParse sample supports one or more of the following task types:

1. **Molecular Image Captioning**  
   Convert molecular diagrams into structured chemical strings.

2. **Symbol and Topology Recognition**  
   Identify atoms, bonds, rings, and connection patterns.

3. **Complex Structure Parsing**  
   Handle abstract rings, variable groups, and non-canonical layouts.

4. **Noise-Robust Recognition**  
   Maintain parsing accuracy under visual distortion or interference.

---

## Data Usage

MolParse is suitable for:
- Training end-to-end optical chemical structure recognition models
- Evaluating vision-language and vision-only chemical parsers
- Scientific document understanding pipelines
- Downstream chemical reasoning and information extraction

---

## Download MolParse Dataset

You can load the MolParse dataset using the HuggingFace `datasets` library:

```python
from datasets import load_dataset

dataset = load_dataset("InnovatorLab/MolParse")
```
## Evaluations

We use [lmms-eval](https://github.com/EvolvingLMMs-Lab/lmms-eval) for evaluations.  
Please refer to the files under [`./evaluation`](./evaluation/README.md) for detailed evaluation configurations and scripts.

---

## License

MolParse 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}
}
```