ThakiOCR-JA-v1
Japanese OCR adapter
2 min read
This model is a research release for recognizing horizontal printed Japanese text lines. It is built on PaddlePaddle/PaddleOCR-VL-1.6 (0.9B) with a merged LoRA adapter trained for 600 steps. The solution targets developers and researchers working with Japanese documents and comes with open-source code for inference. The model is optimized for specific text lines but requires manual image orientation input.
Key features
- Specialized in horizontal printed Japanese text using a LoRA adapter.
- Tested on 2,000 lines from 30 Japanese government white papers.
- Includes vllm library support and an inference script with character normalization.
- Requires explicit image orientation from the calling code as it lacks automatic orientation detection.
What you can do
- Recognize individual printed Japanese text lines in Python scripts.
- Integrate the model into document processing pipelines based on vllm.
- Apply an optional modern_ja normalization mode to fix rare variant glyphs (for example, replacing 齡 with 齢).
Example
from huggingface_hub import snapshot_download
from inference import ThakiOCRJA
model_dir = snapshot_download("ThakiCloud/ThakiOCR-JA-v1")
ocr = ThakiOCRJA(model_dir)
res = ocr.recognize(["line.png"], kind="line", orientation="upright")
print(res[0]["raw_text"])
Read more
DetailsBenchmarks: CER 0.0094 (compared to 0.0118 for the base), Digit F1 0.9973.
Requirements: vllm (version 0.28.0), pillow, model size around 0.96 billion parameters.
Limitations: text lines only, no orientation detection, does not support Korean and English languages.
License: Apache-2.0.










Comments