TextDART-123K: A Dataset for Training and Evaluating TextDART

Project Page

This repository contains TextDART-123K, the training and evaluation dataset introduced in Blame Errant Words, Not the Canvas: Defect-Aware Spatial Credit Assignment for Visual Text Generation. It contains generated images paired with region-level visual text annotations for training TextDART to localize, transcribe, and identify structurally malformed text.

The dataset contains 123K examples: 122K training examples and 1K test examples, with 525K annotated text regions in total.

Due to company policy, some data used in the original research cannot be released. The public dataset has therefore been adjusted and supplemented with replacement samples so that its total number of examples matches the dataset size reported in the original paper.

Code and Project Page

The official implementation, model, training instructions, and project details are available in the WeChatCV/TextDART-SCANNER repository. The released model is available at SujieHu/TextDART-8B-Qwen3VL.

Dataset Format

Each example contains:

• id: the unique example identifier
• images: one image embedded in the Parquet file
• conversations: system, user, and assistant messages used for multimodal SFT

The assistant message follows this format:

{
  "regions": [
    {
      "quad": [x1, y1, x2, y2, x3, y3, x4, y4],
      "recognized_text": "recognized text"
    }
  ]
}

Quadrilateral coordinates are normalized to a 0–1000 grid. Structural anomalies are marked with <#> for one unreadable character, <###> for one unreadable English word, and <#####> for an entirely unreadable block of secondary or fine-print text.

Sample Usage

Load TextDART-123K with the Hugging Face datasets library:

from datasets import load_dataset

# Load both train and test splits.
dataset = load_dataset("SujieHu/TextDART-123K", "default")
train_data = dataset["train"]
test_data = dataset["test"]

# Load one split directly.
train_data = load_dataset("SujieHu/TextDART-123K", "default", split="train")
test_data = load_dataset("SujieHu/TextDART-123K", "default", split="test")

The images are embedded in the Parquet shards, so the complete dataset is large. Use streaming when a local materialized copy is not required:

train_stream = load_dataset(
    "SujieHu/TextDART-123K",
    "default",
    split="train",
    streaming=True,
)

For model training and dataset preparation instructions, see the TextDART training guide.

Citation

If you find this dataset useful for your research, please cite the accompanying paper:

@article{hu2026textdartscanner,
  title={Blame Errant Words, Not the Canvas: Defect-Aware Spatial Credit Assignment for Visual Text Generation},
  author={Hu, Sujie and Zhang, Huaisong and Yang, Pujing and Wang, Qian and Li, Xiaomin and Leng, Xinhang and Li, Chen and Lyu, Jing and Li, Xiu and Pan, Ling},
  journal={arXiv preprint arXiv:2610.20794},
  year={2026}
}