Qwen3.5-35B-A3B-LaTeX-OCR
A fine-tuned version of Qwen/Qwen3.5-35B-A3B for LaTeX OCR tasks, specialized in recognizing handwritten mathematical expressions.
Model Details
Model Description
- Developed by: FutureMa
- Model type: Vision-Language Model (Multimodal)
- Architecture: Mixture of Experts (MoE) - 256 experts, 8 active per token
- Parameters: 35B total, ~3B active
- Fine-tuning Method: LoRA (rank=8, alpha=32, target=all-linear)
- Language(s): English, Chinese
- License: Qwen License
- Finetuned from: Qwen/Qwen3.5-35B-A3B
Uses
Direct Use
This model is designed for:
- Converting handwritten mathematical expressions to LaTeX code
- OCR of mathematical formulas and equations
- Educational and research applications
Out-of-Scope Use
- General image captioning (not optimized for this)
- Non-mathematical handwriting recognition
How to Get Started with the Model
With ms-swift
CUDA_VISIBLE_DEVICES=0,1,2,3 \
swift infer \
--model FutureMa/Qwen3.5-35B-A3B-LaTeX-OCR \
--use_hf true \
--stream true \
--experts_impl grouped_mm \
--device_map auto \
--max_new_tokens 2048
With Transformers
from transformers import AutoModelForCausalLM, AutoProcessor
model = AutoModelForCausalLM.from_pretrained(
"FutureMa/Qwen3.5-35B-A3B-LaTeX-OCR",
torch_dtype="bfloat16",
device_map="auto",
trust_remote_code=True
)
processor = AutoProcessor.from_pretrained(
"FutureMa/Qwen3.5-35B-A3B-LaTeX-OCR",
trust_remote_code=True
)
# Example usage with an image
messages = [
{
"role": "user",
"content": [
{"type": "image", "image": "path/to/math_image.png"},
{"type": "text", "text": "Convert this handwritten math to LaTeX."}
]
}
]
inputs = processor(messages, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512)
result = processor.decode(outputs[0], skip_special_tokens=True)
print(result)
Training Details
Training Data
- Dataset: linxy/LaTeX_OCR (human_handwrite subset)
- Samples: 2,000
Training Procedure
Training Hyperparameters
- Precision: bfloat16
- Learning Rate: 1e-4
- LR Scheduler: Cosine
- Warmup Ratio: 0.05
- Batch Size: 1 (per device)
- Gradient Accumulation Steps: 4
- Epochs: 1
- Max Sequence Length: 2048
- Optimizer: AdamW (fused)
- Weight Decay: 0.1
- LoRA Rank: 8
- LoRA Alpha: 32
Hardware
- GPUs: 4x NVIDIA RTX PRO 6000 (96 GiB VRAM total)
- Training Mode:
device_map auto(single-process multi-GPU, no DeepSpeed)
Training Results
| Metric | Value |
|---|---|
| Final Train Loss | ~0.001 |
| Final Token Accuracy | ~100% |
| Eval Loss | ~0.0006 |
Technical Specifications
Compute Infrastructure
Software
- Framework: ms-swift v4.x
- Transformers: 5.2.0+
- Dependencies: qwen_vl_utils>=0.0.14, decord, torchvision
Bias, Risks, and Limitations
- Fine-tuned on a relatively small dataset (2,000 samples), may not generalize well to all mathematical notation styles
- May struggle with complex or unusual mathematical notations not present in training data
- Requires significant GPU memory for inference (~30GB+ recommended)
- Inherits any biases present in the base Qwen3.5-35B-A3B model
Citation
@misc{qwen3.5-latex-ocr-2026,
author = {FutureMa},
title = {Qwen3.5-35B-A3B-LaTeX-OCR: Fine-tuned Qwen3.5 for Handwritten LaTeX OCR},
year = {2026},
publisher = {Hugging Face},
url = {https://huggingface.co/FutureMa/Qwen3.5-35B-A3B-LaTeX-OCR}
}
Model Card Contact
For questions or issues, please open an issue on the model repository.