abzoo/paddleocr-vl-curriculum-015-lora

PaddleOCR-VL Egyptian ID OCR (Curriculum Learning)

Fine-tuned PaddleOCR-VL for Egyptian national ID OCR using curriculum learning.

Training

Stage: 1/3 Total Epochs: 2 Best Accuracy: 41.2%

Curriculum Structure

1. Pre-train on synthetic Arabic names until saturation
2. Curriculum: real→synthetic→real transition
3. Polish on real data with lower LR

Datasets

• Synthetic: abzoo/arabic-names-synthetic-ocr (2,363 train / 262 val)
• Real: abzoo/egyptian-id-ocr (2,436 train / 211 test)

Results

Accuracy

Metric	Score
Strict	80/211 (37.9%)
Normalized	87/211 (41.2%)

Edit Distance Breakdown

Distance	Count	%
✓ Exact (0)	87	41.2%
~ Close (1)	41	19.4%
~ Close (2)	11	5.2%
~ Close (3)	4	1.9%
✗ Wrong (4+)	68	32.2%

Training History

Epoch	Stage	Strict	Normalized
1	1	31.3%	35.1%
2	1	37.9%	41.2%

Usage

from unsloth import FastVisionModel

model, tokenizer = FastVisionModel.from_pretrained(
    "abzoo/paddleocr-vl-curriculum-015",
    load_in_4bit=True,
)

Base Model

• unsloth/PaddleOCR-VL
• LoRA r=64, alpha=64