Hugging Face's logo

Datasets:
ATH-MaaS
/
TextRenderBench-ZH

Tasks:
Image Classification
Modalities:
Image
Text
Formats:
parquet
Languages:
Chinese
Size:
1K - 10K
ArXiv:
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
TextRenderBench-ZH
File size: 4,261 Bytes
13c361b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
---
language:
- zh
license: apache-2.0
task_categories:
- image-classification
pretty_name: TextRenderBench-ZH
size_categories:
- 1K<n<10K
configs:
- config_name: images
  data_files:
  - split: test
    path: data/images/test-*.parquet
- config_name: annotations
  data_files:
  - split: test
    path: data/annotations/test-*.parquet
---

# TextRenderBench-ZH

TextRenderBench-ZH is an evaluation-only benchmark for target-conditioned verification of structural Chinese glyph rendering. Text-to-image models increasingly produce Chinese text that is readable yet locally wrong, and recognition-oriented systems are invariant to exactly those deviations. This benchmark labels each character crop with its position, intended character, and human-confirmed structural correctness, so detectors can be compared on a shared per-character protocol.

The `test` split holds 299 anonymous source images and 7,127 character-level annotations, of which 608 are rendering errors. There is no training split.

## Resources

| Resource | Location |
|---|---|
| Paper | [TextVerifier: Seeing Beyond Recognition to Verify Chinese Text Rendering](https://arxiv.org/abs/2610.19551) (arXiv:2610.19551) |
| Code | [github.com/ATH-MaaS/TextVerifier](https://github.com/ATH-MaaS/TextVerifier) |
| Model weights | [ATH-MaaS/TextVerifier-CJK-CLIP](https://huggingface.co/ATH-MaaS/TextVerifier-CJK-CLIP) |
| This dataset | [ATH-MaaS/TextRenderBench-ZH](https://huggingface.co/datasets/ATH-MaaS/TextRenderBench-ZH) |

## Dataset configurations

### `images`

The `test` split contains 299 rows:

- `image_id`: anonymous image identifier.
- `image`: embedded RGB PNG image.
- `image_path`: canonical relative path used by the TextVerifier evaluator.
- `width`, `height`: image dimensions; every image is 1024 × 1024.

### `annotations`

The `test` split contains 7,127 rows:

- `sample_id`: unique character-level sample identifier.
- `image_id`: foreign key into the `images` configuration.
- `image_path`: canonical relative image path.
- `item_index`: annotation index within the source image.
- `char`: intended Chinese character.
- `box`: four quadrilateral corner points.
- `rect`: axis-aligned `[x0, y0, x1, y1]` crop rectangle.
- `render_error`: human-confirmed structural rendering-error label.

Join the configurations on `image_id`.

## Labeling protocol

Annotations were initialized from automatic detection and recognition results, then reviewed by human annotators for box placement, character identity, and glyph correctness. A stroke- or component-level deviation is labeled `render_error`; differences of font, calligraphic style, color, and background are not errors on their own.

## Loading

```python
from datasets import load_dataset

images = load_dataset(
    "ATH-MaaS/TextRenderBench-ZH",
    "images",
    split="test",
)
annotations = load_dataset(
    "ATH-MaaS/TextRenderBench-ZH",
    "annotations",
    split="test",
)
```

For reproducible evaluation, pin the Dataset repository to an immutable commit revision with the `revision` argument.

## TextVerifier evaluation

The TextVerifier code repository provides `scripts/materialize_benchmark.py`, which converts these two Parquet configurations into the local PNG and JSONL layout expected by its evaluator.

The evaluation crop rule is:

```python
crop = image.crop((int(x0), int(y0), int(round(x1)), int(round(y1))))
```

## Intended use and limitations

Use this benchmark to evaluate whether a rendered Chinese glyph has a structural error conditioned on its intended character. It is an evaluation set, not a training split. The labels are imbalanced, so report precision, recall, and F1 rather than accuracy alone. Image identifiers are anonymous and must not be interpreted as source identifiers.

## Citation

```bibtex
@article{chen2026textverifier,
  title   = {TextVerifier: Seeing Beyond Recognition to Verify Chinese Text Rendering},
  author  = {Chen, Yiyang and Wang, Guo-Hua and Liu, Yaoli and Chen, Qing-Guo and Xu, Zhao and Luo, Weihua and Ma, Xiujun},
  journal = {arXiv preprint arXiv:2610.19551},
  year    = {2026},
  url     = {https://arxiv.org/abs/2610.19551}
}
```

## License

The benchmark images and annotations are released under the Apache License 2.0. See `LICENSE`.