handwriting-v1

This dataset contains handwriting stroke data collected using a stylus (S Pen) on a tablet device. Optimized for training RNNs (Recurrent Neural Networks) on handwriting generation/recognition tasks.

Data Format

Data is available in two formats in the data/ directory:

• Parquet files (*.parquet): Columnar format, optimized for HuggingFace datasets
• JSONL files (*.jsonl): Line-delimited JSON backup, easy to parse

Both formats contain identical RNN training data with the same batch IDs.

Parquet Schema

Each row in the Parquet files represents a complete handwriting sample:

Column	Type	Description
id	string	Unique identifier (UUID)
text	string	The prompt text that was written
dx	list	Delta X offsets between consecutive points
dy	list	Delta Y offsets between consecutive points
eos	list	End-of-stroke flags (1 = pen lift, 0 = continue)
scale	double	Scale factor used for normalization
created_at	string	ISO timestamp of creation
session_id	string	Collection session identifier

JSONL Format

Each line in the JSONL files is a JSON object with the following structure:

{"id": "uuid", "text": "prompt text", "points": [{"dx": 0, "dy": 0, "eos": 0}, ...], "scale": 1.0}

Field	Type	Description
id	string	Unique identifier (UUID)
text	string	The prompt text that was written
points	array	Array of point objects with dx, dy, eos
scale	number (optional)	Scale factor used for normalization

RNN Training Format

The stroke data is stored in the format commonly used for RNN handwriting models:

• dx/dy: Position deltas from the previous point (first point has dx=dy=0)
• eos: Binary flag indicating pen lifts (end of stroke)
• Data is normalized by bounding box for consistent scale

Visualization

Preview SVGs are available in renders_preview/ for HuggingFace Dataset Viewer.

Usage

Using Parquet (Recommended for HuggingFace)

from datasets import load_dataset

# For private repos, use: load_dataset("finnbusse/handwriting-v1", token="YOUR_HF_TOKEN")
dataset = load_dataset("finnbusse/handwriting-v1")

# Access a sample
sample = dataset['train'][0]

# Stroke data is already native Python lists (no JSON parsing needed)
dx = sample['dx']
dy = sample['dy']
eos = sample['eos']

# Reconstruct absolute positions
x, y = 0, 0
positions = []
for dx_i, dy_i, eos_i in zip(dx, dy, eos):
    x += dx_i
    y += dy_i
    positions.append((x, y, eos_i))

Using JSONL (Alternative)

JSONL filenames follow the batch ID pattern: YYYYMMDD_HHMMSS_XXXX.jsonl

import json
import glob

# Read all JSONL files in the data directory
for jsonl_file in glob.glob('data/*.jsonl'):
    with open(jsonl_file, 'r') as f:
        for line in f:
            sample = json.loads(line)
            points = sample['points']
            scale = sample.get('scale', 1.0)  # scale is optional
            # Each point has: dx, dy, eos

Collection Method

Data was collected using a web application with Pointer Events API, capturing stylus input including pressure and tilt when available.