---
language:
- he
license: cc-by-4.0
task_categories:
- image-to-text
tags:
- ocr
- htr
- handwriting-recognition
- hebrew
- synthetic
size_categories:
- 1M<n<10M
pretty_name: Hebrew Synthetic Handwriting Lines (1M)
configs:
- config_name: default
  data_files:
  - split: train
    path: train-*.parquet
  - split: test
    path: test-*.parquet
---

# Hebrew Synthetic Handwriting Lines — 1M

One million synthetic Hebrew handwritten **text-line** images with transcripts,
built for pre-training Handwriting Text Recognition (HTR / OCR) models such as
TrOCR. Each line is rendered from a large corpus of modern Hebrew sentences using
57 handwriting fonts, with per-line style variation and one of four image
distortions.

## Contents

| | |
|---|---|
| Total images | **1,000,000** |
| Train / Test | **979,869 / 20,131** |
| Unique sentences | 268,034 |
| Image height | 64 px (variable width) |
| Format | RGB PNG, black ink on white |

Each of the four distortion tracks has exactly **250,000** images:

| `distortion` | Effect |
|---|---|
| `clear` | clean render, no image distortion |
| `blur` | Gaussian blur, radius 0.8–1.8 (soft focus / ink bleed) |
| `noise` | additive Gaussian noise, σ 12–35 (scan grain) |
| `rotate` | whole-line rotation ±1–4°, random direction (skewed scan) |

### Split policy — leak-free by sentence

The train/test split is done **by sentence**: all four distortion variants of a
given sentence land in the same split. A naive per-line split would leak a
sentence's `clear` copy into train and its `blurred` copy into test, inflating
apparent accuracy. Test fraction ≈ 2 %.

## Columns

| column | type | description |
|---|---|---|
| `image` | `Image` | the line image (64 px height) |
| `text` | `string` | the ground-truth transcript |
| `distortion` | `string` | one of `clear` / `blur` / `noise` / `rotate` |

## Usage

```python
from datasets import load_dataset

ds = load_dataset("cyttic/heb-synth-1m")           # train + test
print(ds["train"][0]["text"], ds["train"][0]["image"].size)

# e.g. only the clean track:
clear = ds["train"].filter(lambda r: r["distortion"] == "clear")
```

## How it was generated

- **Text.** ~268k modern Hebrew sentences generated by an LLM
  (**DictaLM 2.0 Instruct**), seed-conditioned on mid/low-frequency Hebrew words
  from a Wikipedia word-frequency list, filtered to Hebrew-only and ≤ ~12 words,
  exact-deduplicated. This deliberately covers **modern** vocabulary that
  literary Hebrew corpora miss.
- **Rendering.** 57 handwriting fonts (two families). Each line uses a random
  font plus randomized style knobs — letter-shape jitter, letter spacing
  ("tightness"), up/down baseline wobble, and small per-letter rotation — so the
  same sentence never renders identically. Malformed font glyphs (blobs, stray
  dirt-dots, broken strokes) are auto-detected and repaired by synthesizing a
  clean glyph, so the corpus is clean across all fonts.
- **Distortion.** After rendering, one of the four distortions above is applied.
- **Constraint.** Lines are kept within the 2304 px width limit (at 64 px height)
  used by block-style ViT preprocessing.

## Intended use & limitations

Intended for **Stage-1 synthetic pre-training** of Hebrew HTR models, followed by
fine-tuning on real handwritten data. It is synthetic: fonts approximate but do
not replace genuine human handwriting variability (pen pressure, real paper
backgrounds, connected cursive). No paper-texture backgrounds are included —
images are on clean white.

## Citation / provenance

Synthetic dataset generated by the DiffPen-Hebrew glyph-synthesis pipeline. Fonts
are Hebrew handwriting fonts; verify individual font licenses before commercial
use.