README.md

---
license: apache-2.0
language:
- en
task_categories:
- image-to-text
tags:
- ocr
- synthetic
- vision
- trocr
- llava
- florence-2
- text-recognition
- computer-vision
size_categories:
- 100K<n<1M
dataset_info:
  features:
  - name: image
    dtype: image
  - name: text
    dtype: string
  splits:
  - name: train
    num_bytes: 556311460.0
    num_examples: 100000
  download_size: 557311624
  dataset_size: 556311460.0
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
---

# 👁️ Legend-OCR Dataset

Welcome to the **Legend-OCR** dataset! This is a highly robust, synthetically generated Vision dataset designed specifically for training **Optical Character Recognition (OCR)** models and **Vision-Language Models (VLMs)** like TrOCR, Florence-2, LLaVA, and Qwen-VL.

## 📊 Dataset Overview
- **Name:** Legend-OCR
- **Type:** Synthetic Vision-Text Pair
- **Size:** 100,000 High-Quality Images (Configurable)
- **Format:** Parquet (Embedded PNG Bytes + String Text)
- **Task:** Image-to-Text / OCR
- **Characters Covered:** Alphabets (A-Z, a-z), Numbers (0-9), and all standard Punctuation/Symbols.

---

## 🚀 Key Features & Generation Logic

This dataset was procedurally generated using Python (PIL) with advanced randomization techniques to make the AI models robust against real-world variations:

1. **Massive Font Variety:** Uses multiple Linux-native fonts (`Ubuntu`, `Roboto`, `Noto`, `Liberation`) encompassing Regular, Bold, Italic, and Thin styles.
2. **Dynamic Text Lengths:**
   - 30% of the dataset features **Single Characters** (perfect for basic symbol recognition and bounding box training).
   - 70% features **10 to 20 Characters** (perfect for word and sentence-level context recognition).
3. **High-Contrast Backgrounds:**
   - 50% Images: Dark Background with Light/White Text.
   - 50% Images: Light Background with Dark/Black Text.
4. **Dynamic Image Sizing:** Bounding boxes and image canvas sizes scale automatically based on text length and randomized padding, teaching the model to focus on the subject rather than a fixed aspect ratio.
5. **Zero Hallucination:** Since the dataset is synthetically generated natively in code, the ground truth text has a **100% accuracy rate**.

---

## 📂 Dataset Structure

Under the hood, the dataset is saved in highly compressed Parquet format. The schema looks like this:

```json
{
  "image": {
    "bytes": "\u0089PNG\r\n\u001a\n\u0000\u0000\u0000\rIHDR...",
    "path": null
  },
  "text": "Hello@123!"
}