OsGo

Qwen3.5-ocr-jp-2b

Deploy Dedicated

README

License: apache-2.0

Focus

Training data emphasizes the following Japanese document features:

Ruby annotations — emitted as HTML5 ruby markup, e.g. <ruby>漢字<rt>かんじ</rt></ruby>
Japanese handwriting, vertical writing

Quickstart

vLLM (recommended)

python

import base64, iofrom PIL import Imagefrom vllm import LLM, SamplingParams PROMPT = "OCR this image as HTML layout blocks with bbox and label." llm = LLM(    model="ebinan92/Qwen3.5-ocr-jp-2b",    dtype="bfloat16",    max_model_len=12288,    limit_mm_per_prompt={"image": 1},    trust_remote_code=True,)sampling = SamplingParams(temperature=0.0, top_p=0.1, max_tokens=8000) image = Image.open("page.png").convert("RGB")buf = io.BytesIO()image.save(buf, format="PNG")b64 = base64.b64encode(buf.getvalue()).decode() messages = [{    "role": "user",    "content": [        {"type": "image_url", "image_url": {"url": f"data:image/png;base64,{b64}"}},        {"type": "text", "text": PROMPT},    ],}]print(llm.chat(messages, sampling_params=sampling)[0].outputs[0].text)

Requires vllm>=0.19.1 and transformers>=5.5.1.

transformers

python

import torchfrom PIL import Imagefrom transformers import AutoProcessor, AutoModelForImageTextToText PROMPT = "OCR this image as HTML layout blocks with bbox and label." ckpt = "ebinan92/Qwen3.5-ocr-jp-2b"processor = AutoProcessor.from_pretrained(ckpt)model = AutoModelForImageTextToText.from_pretrained(    ckpt, dtype=torch.bfloat16, device_map="auto") image = Image.open("page.png").convert("RGB")messages = [{    "role": "user",    "content": [        {"type": "image", "image": image},        {"type": "text", "text": PROMPT},    ],}]inputs = processor.apply_chat_template(    messages,    add_generation_prompt=True,    tokenize=True,    return_dict=True,    return_tensors="pt",).to(model.device) out = model.generate(**inputs, max_new_tokens=8000, do_sample=False)print(processor.batch_decode(out[:, inputs["input_ids"].shape[1]:], skip_special_tokens=True)[0])

Benchmarks

Table with columns: Benchmark, Metric, chandra-ocr-2, Qwen3.5-ocr-jp-2b, sarashina2.2-ocr
Benchmark	Metric	chandra-ocr-2	Qwen3.5-ocr-jp-2b	sarashina2.2-ocr
olmOCR-bench	Accuracy ↑	85.9†	82.8	—
VJRODa※	CER % ↓	7.2	7.3	12.0
※

sarashina2.2-ocr's olmOCR-bench overall is omitted because its HF card does not report the baseline row.

※ VJRODa is evaluated on 92 / 100 samples (8 PDFs are NDL WARP-restricted and unavailable).
† olmOCR-bench score for chandra-ocr-2 is taken from the official HF card.

Table with columns: JSONL, chandra-ocr-2†, Qwen3.5-ocr-jp-2b
JSONL	chandra-ocr-2†	Qwen3.5-ocr-jp-2b
arxiv_math	90.2	85.7
table_tests	89.9	88.1
baseline	99.6	99.1
headers_footers	92.5	90.3
old_scans_math	89.3	81.9
long_tiny_text

Limitations

Works only with the single fixed prompt above. It is not tuned for other tasks or free-form instructions.
Trained primarily on Japanese and English. Coverage of other languages (Chinese, Korean, etc.) is incidental.

License

Apache 2.0.

This model is derived from Qwen3.5-2B, trained on independently constructed datasets. No outputs or weights from datalab-to/chandra-ocr-2 (or any other Chandra release) were used.

Acknowledgements

Qwen3.5-2B — base model (Apache 2.0)
Chandra — format reference

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

OsGo

Model Tree

Base

Qwen/Qwen3.5-2B

Fine-tuned

this model

Input Modalities

TextImageVideo

Output Modalities

Text

Supported Functionality

Dedicated Endpoints

Explore FriendliAI today

Get started Talk to an engineer

README

License: apache-2.0

Focus

Training data emphasizes the following Japanese document features:

Ruby annotations — emitted as HTML5 ruby markup, e.g. <ruby>漢字<rt>かんじ</rt></ruby>
Japanese handwriting, vertical writing

Quickstart

vLLM (recommended)

python

import base64, iofrom PIL import Imagefrom vllm import LLM, SamplingParams PROMPT = "OCR this image as HTML layout blocks with bbox and label." llm = LLM(    model="ebinan92/Qwen3.5-ocr-jp-2b",    dtype="bfloat16",    max_model_len=12288,    limit_mm_per_prompt={"image": 1},    trust_remote_code=True,)sampling = SamplingParams(temperature=0.0, top_p=0.1, max_tokens=8000) image = Image.open("page.png").convert("RGB")buf = io.BytesIO()image.save(buf, format="PNG")b64 = base64.b64encode(buf.getvalue()).decode() messages = [{    "role": "user",    "content": [        {"type": "image_url", "image_url": {"url": f"data:image/png;base64,{b64}"}},        {"type": "text", "text": PROMPT},    ],}]print(llm.chat(messages, sampling_params=sampling)[0].outputs[0].text)

Requires vllm>=0.19.1 and transformers>=5.5.1.

transformers

python

import torchfrom PIL import Imagefrom transformers import AutoProcessor, AutoModelForImageTextToText PROMPT = "OCR this image as HTML layout blocks with bbox and label." ckpt = "ebinan92/Qwen3.5-ocr-jp-2b"processor = AutoProcessor.from_pretrained(ckpt)model = AutoModelForImageTextToText.from_pretrained(    ckpt, dtype=torch.bfloat16, device_map="auto") image = Image.open("page.png").convert("RGB")messages = [{    "role": "user",    "content": [        {"type": "image", "image": image},        {"type": "text", "text": PROMPT},    ],}]inputs = processor.apply_chat_template(    messages,    add_generation_prompt=True,    tokenize=True,    return_dict=True,    return_tensors="pt",).to(model.device) out = model.generate(**inputs, max_new_tokens=8000, do_sample=False)print(processor.batch_decode(out[:, inputs["input_ids"].shape[1]:], skip_special_tokens=True)[0])

Benchmarks

Table with columns: Benchmark, Metric, chandra-ocr-2, Qwen3.5-ocr-jp-2b, sarashina2.2-ocr
Benchmark	Metric	chandra-ocr-2	Qwen3.5-ocr-jp-2b	sarashina2.2-ocr
olmOCR-bench	Accuracy ↑	85.9†	82.8	—
VJRODa※	CER % ↓	7.2	7.3	12.0
※

sarashina2.2-ocr's olmOCR-bench overall is omitted because its HF card does not report the baseline row.

※ VJRODa is evaluated on 92 / 100 samples (8 PDFs are NDL WARP-restricted and unavailable).
† olmOCR-bench score for chandra-ocr-2 is taken from the official HF card.

Table with columns: JSONL, chandra-ocr-2†, Qwen3.5-ocr-jp-2b
JSONL	chandra-ocr-2†	Qwen3.5-ocr-jp-2b
arxiv_math	90.2	85.7
table_tests	89.9	88.1
baseline	99.6	99.1
headers_footers	92.5	90.3
old_scans_math	89.3	81.9
long_tiny_text

Limitations

Works only with the single fixed prompt above. It is not tuned for other tasks or free-form instructions.
Trained primarily on Japanese and English. Coverage of other languages (Chinese, Korean, etc.) is incidental.

License

Apache 2.0.

This model is derived from Qwen3.5-2B, trained on independently constructed datasets. No outputs or weights from datalab-to/chandra-ocr-2 (or any other Chandra release) were used.

Acknowledgements

Qwen3.5-2B — base model (Apache 2.0)
Chandra — format reference