Sam4rano/paddleocr-vl15-yoruba-lora

Yorùbá OCR LoRA Adapter (PaddleOCR-VL-1.5)

This is a LoRA adapter for the PaddleOCR-VL-1.5 model, fine-tuned specifically for Yorùbá line crops as part of ongoing OCR research.

Training Details

• —Dataset: ~2,300 unique Yorùbá text line images.
• —Epochs: 5
• —Architecture: LoRA (r=16) on Qwen2-VL visual/text projection layers.
• —Normalization: Unicode NFC normalization with Yorùbá combining diacritics.

Research Results (Table 1 Comparison)

Usage

To use this adapter, load the base model using transformers and apply the adapter using peft.