End-to-end vision-language OCR for pre-modern and early-modern Japanese documents. Full fine-tune of Qwen/Qwen3.5-2B on a combined corpus of printed honkoku transcriptions and Kuzushiji cursive pages.

Training data

NDL minhon-v2 honkoku (~20.4k pages) — printed Japanese transcriptions across 15 categories (tempo, ryukyu, gakugei, dafukui, kusazoushi, …), with cleaning applied (parenthetical furigana stripped, kanbun markers stripped, brackets stripped, per-category length filtering).
Kaggle Kuzushiji recognition corpus (~3.5k pages) — cursive Japanese pages held back from any evaluation split.

Total: ~23.9k training pages. Held-out eval set: 416 honkoku rows (cleaned_v2 val) + 120 Kuzushiji rows (kaggle held-out val).

Quick start (vLLM)

bash

vllm serve tohoku-cijs/qwen3.5-2b-ndl-honkoku-kuzushiji-ocr \
    --served-model-name ocr --trust-remote-code \
    --max-model-len 16384 --gpu-memory-utilization 0.7

python

import base64, requests

with open("page.jpg", "rb") as f:
    b64 = base64.b64encode(f.read()).decode()

r = requests.post(
    "http://localhost:8000/v1/chat/completions",
    json={
        "model": "ocr",
        "messages": [{
            "role": "user",
            "content": [{"type": "image_url",
                         "image_url": {"url": f"data:image/jpeg;base64,{b64}"}}],
        }],
        "temperature": 0.0,
        "max_tokens": 2048,
    },
)
print(r.json()["choices"][0]["message"]["content"])

License & attribution

Base model: Qwen3.5-2B — see Qwen license.
Training data: NDL minhon-v2 (CC BY 4.0) and Kaggle Kuzushiji recognition corpus (CC BY-SA 4.0). Cite the underlying sources if you publish results.