AntoniusPerf/runic-ocr-qwen-vl-lora

RUNIC-OCR — QLoRA adapters for runic inscription recognition

LoRA adapters (QLoRA: 4-bit NF4 base + LoRA r=16, α=32, dropout 0.05) from the master's thesis «Автоматическое распознавание, перевод и анализ древнегерманских рунических текстов» (A. Perfilev, HSE University, 2026).

Task: photo of a runic inscription → Latin transliteration in the Rundata convention. Trained only on synthetic images (SD3 + ControlNet Canny, ~4.6k images), evaluated on a real gold set of 113 lines.

subfolder	base model	CER synth val, %	CER gold, % (95% CI)
qwen25vl-7b/	Qwen/Qwen2.5-VL-7B-Instruct	12.03	54.27 (48.0–61.3)
qwen3vl-8b/	Qwen/Qwen3-VL-8B-Instruct	34.69	67.86 (62.2–73.6)
qwen3vl-2b/	Qwen/Qwen3-VL-2B-Instruct	33.39	73.55 (69.9–77.5)

Code, data and the thesis: https://github.com/kekys778/RUNIC-OCR

from peft import PeftModel
from transformers import AutoProcessor, Qwen2_5_VLForConditionalGeneration
base = Qwen2_5_VLForConditionalGeneration.from_pretrained("Qwen/Qwen2.5-VL-7B-Instruct", device_map="auto", load_in_4bit=True)
model = PeftModel.from_pretrained(base, "AntoniusPerf/runic-ocr-qwen-vl-lora", subfolder="qwen25vl-7b")
processor = AutoProcessor.from_pretrained("AntoniusPerf/runic-ocr-qwen-vl-lora", subfolder="qwen25vl-7b")