KADİM OCR LoRA (kadim-ocr-lora)

A LoRA adapter for Qwen2.5-VL-7B-Instruct that reads Turkish document images and returns their text. It cuts the character error rate from 0.428 to 0.108 on a held-out set of degraded scans, with the largest gains on Turkish-specific letters (ı, ş, ğ, ö, ü, ç).

It is the reader (OCR) component of KADİM — Kamu Dilekçe ve İnceleme Mekanizması, an end-to-end agent system for public-sector correspondence built for the TEKNOFEST 2026 Language Agents Competition (Scenario 1).

• Project code: github.com/furkancmc/teknofest-kadim-kamu-evrak-ajani (QWEN_VL_ocr/)
• Sibling adapters: kadim-analiz-lora, kadim-yazi-lora

Model details

Base model	Qwen/Qwen2.5-VL-7B-Instruct
Method	LoRA in bf16; the vision encoder is frozen and adapters are added to the language-model layers only
LoRA rank / alpha / dropout	16 / 32 / 0.05
Language	Turkish
License	MIT (the base model is Apache-2.0)

Note: adapter_config.json stores the training-time local path of the base model. Load the base from the Hub as shown below; PEFT uses the base you pass in.

Intended use

Turning scanned or photographed Turkish petitions and official letters into plain text, as the first stage of the KADİM pipeline (OCR → analysis → legislation retrieval → officer decision → letter draft).

How to use

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

base_id = "Qwen/Qwen2.5-VL-7B-Instruct"
adapter_id = "furkancmc/kadim-ocr-lora"

processor = AutoProcessor.from_pretrained(base_id)
model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
    base_id, torch_dtype=torch.bfloat16, device_map="auto"
)
model = PeftModel.from_pretrained(model, adapter_id)

image = Image.open("document.png").convert("RGB")
messages = [{
    "role": "user",
    "content": [
        {"type": "image"},
        {"type": "text", "text": "Bu belgedeki metni olduğu gibi yaz."},  # use the prompt from the training notebook
    ],
}]
prompt = processor.apply_chat_template(messages, add_generation_prompt=True)
inputs = processor(text=[prompt], images=[image], return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=2048, do_sample=False)
print(processor.batch_decode(out[:, inputs.input_ids.shape[1]:], skip_special_tokens=True)[0])

For the exact instruction text used in training and evaluation, see QWEN_VL_ocr/qwen_ocr_finetune_lora.ipynb. In the full system the adapter is served with vLLM next to the Qwen2.5-VL base.

Training data

Dataset: furkancmc/kadim-ocr-data (train, validation and test splits).

1,000 synthetic document images (800 train / 100 validation / 100 test). The text comes from the open erdem-erdem/Turkish-Law-Documents-700k-clustered corpus of public court decisions (Yargıtay and Danıştay). generate_dataset_1000.py re-typesets each text as a letterhead page and degrades it to simulate scans, in four equal groups of 250:

Group	Scan quality
A	Clean print, no degradation
B	Slight blur, small contrast shift
C	Clear blur, reduced contrast, noise
D	Low resolution, heavy noise, skew

The images are synthetic training pages, not scans of real official documents, and contain no personal data of real citizens.

Evaluation

Test set: 100 document images, 21,571 Turkish characters. Base model and adapter were evaluated on the same set under the same conditions.

Model	Character error rate (CER) ↓	Turkish-character accuracy ↑
Qwen2.5-VL-7B-Instruct (base)	0.4283	78.55%
+ this adapter	0.1076	85.90%

CER drops by about 75% (roughly 4×). Notebooks: qwen_ocr_baseline_test.ipynb (base) and qwen_ocr_finetune_lora.ipynb (adapter).

Limitations

• Trained and tested on synthetic degradations of typeset text; real scans, handwriting, stamps and unusual layouts are not covered by the evaluation.
• Turkish-character accuracy is still 85.9%, so OCR output should be treated as noisy input to the downstream analysis step.
• Long multi-page documents should be processed page by page.

Citation

@misc{kadim2026,
  title  = {KADİM: Kamu Dilekçe ve İnceleme Mekanizması},
  author = {Camcıoğlu, Furkan},
  year   = {2026},
  url    = {https://github.com/furkancmc/teknofest-kadim-kamu-evrak-ajani}
}