README.md

---
license: mit
dataset_info:
- config_name: pretrain_synthetic_7M
  features:
  - name: image
    dtype: image
  - name: SMILES
    dtype: string
  splits:
  - name: train
    num_bytes: 115375911760.028
    num_examples: 7720468
  download_size: 122046202421
  dataset_size: 115375911760.028
- config_name: sft_real
  features:
  - name: image
    dtype: image
  - name: SMILES
    dtype: string
  splits:
  - name: train
    num_bytes: 2479379042.298
    num_examples: 91166
  download_size: 2416204649
  dataset_size: 2479379042.298
- config_name: test_markush_10k
  features:
  - name: image
    dtype: image
  - name: SMILES
    dtype: string
  splits:
  - name: train
    num_bytes: 228019568
    num_examples: 10000
  download_size: 233407872
  dataset_size: 228019568
- config_name: test_simple_10k
  features:
  - name: image
    dtype: image
  - name: SMILES
    dtype: string
  splits:
  - name: train
    num_bytes: 291640094
    num_examples: 10000
  download_size: 292074581
  dataset_size: 291640094
- config_name: valid
  features:
  - name: image
    dtype: image
  - name: SMILES
    dtype: string
  splits:
  - name: train
    num_bytes: 13538058
    num_examples: 403
  download_size: 13451383
  dataset_size: 13538058
configs:
- config_name: pretrain_synthetic_7M
  data_files:
  - split: train
    path: pretrain_synthetic_7M/train-*
- config_name: sft_real
  data_files:
  - split: train
    path: sft_real/train-*
- config_name: test_markush_10k
  data_files:
  - split: train
    path: test_markush_10k/train-*
- config_name: test_simple_10k
  data_files:
  - split: train
    path: test_simple_10k/train-*
- config_name: valid
  data_files:
  - split: train
    path: valid/train-*
tags:
- chemistry
---

# MolParser-7M

**Anonymous Open Source now**

This repo provides the training data and evaluation data for MolParser, proposed in paper *“MolParser: End-to-end Visual Recognition of Molecule Structures in the Wild“* (ICCV2025 under review)

MolParser-7M contains nearly 8 million paired image-SMILES data. It should be noted that the caption of image is our extended-SMILES format, which suggested in our paper.

* **MolParser-Pretrain**: More than 7.7M synthetic training data in `pretrain_synthetic_7M` subset;

* **MolParser-SFT [WIP]**: Nearly 400k samples for fine-tuning stage in `sft_real` subset. (We are organizing an OCSR competition based on MolParser-7M, so we have reserved part of the MolParser-SFT data for the competition. Stay tuned!)

* **MolParser-Val**: A small validation set carefully selected in-the-wild in `valid` subset. It can be used to quickly valid the model ability during the training process;

* **WildMol Benchmark**: 20k molecule structure images cropped from real patents or paper, `test_simple_10k`（WildMol-10k）subset and `test_markush_10k`（WildMol-10k-M）subset;

As the paper is still **under review**, this data is provided **anonymously**. More information will be provided after the paper has been accepted.

In the future, we will continue to re-clean this dataset, open-source more data, update model checkpoints, refresh the benchmark results, and release the training code.

[**Anonymous Demo: Click Here**](http://101.126.35.171:50008/)