Qwen3.5-2B-MathParser-pro

Model Summary

Qwen3.5-2B-MathParser-pro is a compact vision-language model for handwritten mathematical formula OCR. It is optimized to transcribe single-line and multi-line handwritten mathematical expressions into LaTeX, with a focus on local deployment.

This 2B release is intended for lower-memory local deployment. The companion release is Qwen3.5-4B-MathParser-pro.

Intended Use

• Handwritten mathematical formula recognition
• Multi-line LaTeX transcription
• OCR for mathematical expressions and derivations
• Research and application prototyping around handwritten math parsing

This model is not intended to be a general mathematical reasoning model. It should be used as an OCR/transcription model.

Training Recipe

The model follows a two-stage MathParser training recipe:

1. Stage 1 SFT builds a stable handwritten mathematical OCR base and teaches direct LaTeX transcription.
2. Stage 2 DPO v34 prefers concise, stable, line-count-faithful transcriptions and reduces malformed outputs, repetition, max-token runaway, and very low-similarity failures.

The released weights are fully merged model weights, not LoRA adapters.

Evaluation

Evaluation set: 756 multi-line handwritten mathematical formula samples.

Metrics:

• Avg Sim / Median Sim: normalized edit similarity, higher is better.
• Line Match: exact line-count match with ground truth.
• Within +/-1: predicted line count differs from ground truth by at most one.
• Runaway: max-token or obviously overlong/repetitive generations, lower is better.
• Bad <0.50: samples with similarity below 0.50, lower is better.

Model	Samples	Avg Sim	Median Sim	Line Match	Within +/-1	Runaway	Bad <0.50
Qwen3.5-0.8B Base	756	0.544843	0.580742	149	235	108	262
Qwen3.5-2B Base	756	0.599258	0.651649	252	392	19	236
Qwen3.5-4B Base	756	0.534456	0.541674	264	368	5	295
Qwen3.5-2B SFT	756	0.906516	0.952732	550	706	13	25
Qwen3.5-2B SFT+DPO	756	0.916060	0.951464	569	714	3	15
Qwen3.5-4B SFT	756	0.942045	0.966546	612	730	0	2
Qwen3.5-4B SFT+DPO	756	0.942878	0.968560	611	730	0	1

For this release, the main result is:

Release	Avg Sim	Median Sim	Line Match	Within +/-1	Runaway	Bad <0.50
Qwen3.5-2B-MathParser-pro	0.916060	0.951464	569	714	3	15

Figures

Usage

from PIL import Image
import torch
from transformers import AutoModelForImageTextToText, AutoProcessor
from qwen_vl_utils import process_vision_info

model_id = "sugartai/Qwen3.5-2B-MathParser-pro"

processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForImageTextToText.from_pretrained(
    model_id,
    trust_remote_code=True,
    dtype=torch.bfloat16,
    device_map="auto",
).eval()

image = Image.open("formula.png").convert("RGB")
messages = [
    {
        "role": "system",
        "content": "You are a handwritten mathematical OCR model. Return only the LaTeX transcription.",
    },
    {
        "role": "user",
        "content": [
            {"type": "image", "image": image},
            {"type": "text", "text": "Transcribe the handwritten mathematical formula into LaTeX only."},
        ],
    },
]

text = processor.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=False,
)
image_inputs, video_inputs = process_vision_info(messages)
inputs = processor(
    text=[text],
    images=image_inputs,
    videos=video_inputs,
    padding=True,
    return_tensors="pt",
).to(model.device)

eos_ids = [processor.tokenizer.eos_token_id]
pad_id = processor.tokenizer.pad_token_id
if pad_id is not None and pad_id not in eos_ids:
    eos_ids.append(pad_id)

with torch.no_grad():
    output_ids = model.generate(
        **inputs,
        max_new_tokens=1536,
        do_sample=False,
        num_beams=1,
        eos_token_id=eos_ids,
        pad_token_id=pad_id if pad_id is not None else eos_ids[0],
    )

new_ids = output_ids[:, inputs["input_ids"].shape[1]:]
print(processor.decode(new_ids[0], skip_special_tokens=True))

Limitations

• The model is specialized for handwritten mathematical OCR and LaTeX transcription.
• It is not a general reasoning or theorem-proving model.
• Very noisy images, unusual notation, extreme layout variation, or out-of-distribution handwriting may degrade quality.
• The reported metrics are from an internal 756-sample multi-line handwritten formula evaluation set.

License

This model is released under Apache 2.0, following the base model license of Qwen/Qwen3.5-2B.

Citation

If you use this model, please cite or link this model page and the Qwen3.5 base model.