Modelpublic

ModarIbrahim/road-qwen25vl-lora-e3

sourceHugging Faceupdated 8d agoView on Hugging Face

0likes13downloads

Model Card

R.O.A.D. Barbados Historic Handwriting — Qwen2.5-VL-7B LoRA

LoRA adapter for line-level transcription of historic Barbados handwriting (Zindi R.O.A.D. challenge).

—Base: Qwen/Qwen2.5-VL-7B-Instruct
—LoRA: r=32, alpha=64, dropout=0.05, all-linear, bf16
—Trained only on competition data (no external OCR datasets)
—2 epochs, image height 256, max_pixels 28282048
—Held-out val: WER 0.1551 / CER 0.0441 / score 0.9004; Zindi public LB 0.90413

Usage

python

from transformers import AutoProcessor, Qwen2_5_VLForConditionalGeneration
from peft import PeftModel
import torch

base = Qwen2_5_VLForConditionalGeneration.from_pretrained(
    "Qwen/Qwen2.5-VL-7B-Instruct", dtype=torch.bfloat16).to("cuda")
model = PeftModel.from_pretrained(base, "ModarIbrahim/road-qwen25vl-lora").eval()
processor = AutoProcessor.from_pretrained(
    "Qwen/Qwen2.5-VL-7B-Instruct", min_pixels=256*28*28, max_pixels=28*28*2048)
processor.tokenizer.padding_side = "left"