---
license: cc-by-4.0
task_categories:
- image-to-text
language:
- ja
tags:
- ocr
- japanese
- text-recognition
- book-ocr
- scene-text
pretty_name: honmono-ocr test set
size_categories:
- 10K<n<100K
---

# honmono-ocr test set

A held-out benchmark of 14,256 real Japanese text lines, each a cropped line image paired with its transcription. It is the evaluation set for the [honmono-ocr](https://github.com/eridgd/honmono-ocr) recognition models, intended both for reproducing their reported accuracy and for benchmarking other Japanese line-recognition models on the same data.

The lines are drawn from printed books, scanned print, and scene text, spanning the conditions a camera-based Japanese OCR system typically meets.

## Dataset structure

A single `test` split of 14,256 examples. Each example has three fields:

| field | type | description |
|---|---|---|
| `image` | image | one cropped text line |
| `text` | string | transcription, NFKC-normalized |
| `source` | string | origin: `ndl_pdm`, `ndl_ndlocr`, `icdar_mlt`, or `ndl_oneline` |

Reported scores for the honmono-ocr models are computed against the NFKC-normalized `text`.

Composition by source:

| source | lines | description |
|---|---:|---|
| `ndl_pdm` | 8,768 | scanned pre-war printed books (National Diet Library) |
| `icdar_mlt` | 4,114 | Japanese scene text (signs, storefronts, displays) |
| `ndl_ndlocr` | 1,353 | National Diet Library print, more modern |
| `ndl_oneline` | 21 | a small NDL one-line set, reported for completeness |

Both horizontal and vertical (縦書き) lines are included. Vertical crops are stored rotated 90°, the form the models consume at inference.

## Usage

```python
from datasets import load_dataset

ds = load_dataset("eridgd/honmono-ocr-test", split="test")
print(ds[0]["text"], ds[0]["source"])
ds[0]["image"]
```

To reproduce the published honmono-ocr scores, use the evaluation runner in the [code repository](https://github.com/eridgd/honmono-ocr). This dataset bundles `test_real.manifest.txt`, which lists the same 14,256 lines:

```bash
python eval/run_eval.py \
  --model models/v6-small/honmono-v6-small-fp16.onnx \
  --manifest path/to/honmono-ocr-test/test_real.manifest.txt \
  --crops_dir path/to/honmono-ocr-test
```

Reference results, whole-line accuracy (a line counts as correct only when every character matches):

| model | overall | NDL PDM | NDL NDLOCR | ICDAR MLT |
|---|---:|---:|---:|---:|
| honmono v6-small | 83.5 | 87.5 | 83.2 | 75.2 |
| honmono v6-tiny | 73.6 | 80.7 | 77.8 | 57.3 |
| PP-OCRv6-small (off-the-shelf) | 50.2 | 50.8 | 48.9 | 49.6 |
| PP-OCRv6-tiny (off-the-shelf) | 26.2 | 29.8 | 22.3 | 19.8 |

## Source data and licensing

This dataset is redistributed under CC-BY-4.0, the license of both source datasets, and contains material modified from the originals as noted below.

**National Diet Library (NDL) OCR datasets** (`ndl_pdm`, `ndl_ndlocr`, `ndl_oneline`). © National Diet Library, Japan. Licensed CC-BY-4.0. Source: <https://lab.ndl.go.jp/>

**ICDAR Robust Reading Challenge on Multi-Lingual Scene Text — RRC-MLT, 2017 and 2019** (`icdar_mlt`). Licensed CC-BY-4.0. Source: <https://rrc.cvc.uab.es/?ch=8> (2017) and <https://rrc.cvc.uab.es/?ch=15> (2019). Dataset described in N. Nayef et al., *ICDAR2019 Robust Reading Challenge on Multi-Lingual Scene Text Detection and Recognition (RRC-MLT-2019)*, ICDAR 2019.

Modifications applied to both sources: restricted to Japanese text; images cropped to individual text-line regions; annotations converted to a single `text` field and NFKC-normalized; unusable samples removed. RRC-MLT-2019's test-set ground truth was not publicly released and is not included; the ICDAR lines are taken from the publicly released train/validation portions. No endorsement by the original creators is implied.

The NDL portion (approximately 10,000 lines) requires no registration. The ICDAR portion originates from the RRC site, which requires a free account to download the source data.

The ICDAR MLT portion consists of photographs of real-world scenes. CC-BY-4.0 grants the copyright and database rights held by the licensors; it does not address privacy, publicity, or trademark rights that may attach to people, signage, or logos visible in individual images. This is standard for scene-text datasets and does not restrict ordinary research or benchmarking use. Anyone reusing specific images in other contexts should evaluate those rights independently.

## Citation

For the honmono-ocr models and this dataset packaging:

```bibtex
@software{honmono_ocr,
  title  = {honmono-ocr: on-device Japanese book OCR},
  author = {Evan Davis},
  year   = {2026},
  url    = {https://github.com/eridgd/honmono-ocr}
}
```

Users must also credit the source datasets (NDL Lab; ICDAR RRC-MLT) under their CC-BY-4.0 terms.

## Links

- Models: <https://huggingface.co/eridgd/honmono-ocr-v6-small> (also `-v6-tiny`, `-v5`)
- Code, training recipe, and evaluation runner: <https://github.com/eridgd/honmono-ocr>