Qwen2_5_3B_VL_HandWritten_LaTeX_OCR

This repository contains code, dataset setup, and scripts used to fine-tune Qwen2.5-VL-3B for converting handwritten mathematical expressions into LaTeX code.

Ideal for automating LaTeX conversion in:

Math learning apps Academic tools E-notes platforms Personal study or research agents

Model available on Hugging Face: https://huggingface.co/aaghaazkhan/Qwen2_5_3B_VL_HandWritten_LaTeX_OCR

---

Overview

This project demonstrates how to:

• Fine-tune a multimodal vision-language model using LoRA
• Train a model on low VRAM (6GB GPU)
• Convert handwritten equations into clean LaTeX
• Measure performance using Exact Match and Token Accuracy
• Run inference locally with Transformers

---

Features

• Handwritten Math to LaTeX Code
• Trained on NVIDIA RTX 3050 (6GB)
• Total VRAM usage: 5.72 GB
• Qwen 2.5-VL-3B multimodal base
• Efficient fine-tuning using 4-bit quantization + LoRA

---

Base model vs Fine-tuned model outputs:

Base model:

Sample 1:

Sample 2:

Fine-tuned model:

Sample 1:

Sample 2:

---

Training Performance (via WandB)

Training Loss Curve

Validation Loss Curve

---

Inference Code

from transformers import Qwen2_5_VLForConditionalGeneration, AutoTokenizer, AutoProcessor
from qwen_vl_utils import process_vision_info

model_id = "aaghaazkhan/Qwen2_5_3B_VL_HandWritten_LaTeX_OCR"
model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
    model_id, torch_dtype="auto", device_map="auto"
)

processor = AutoProcessor.from_pretrained(model_id)

messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "image",
                "image": "latex_demo.png",
            },
            {"type": "text", "text": "Convert the mathematical content in the image into LaTeX."},
        ],
    }
]

# Preparation for inference
text = processor.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True
)
image_inputs, video_inputs = process_vision_info(messages)
inputs = processor(
    text=[text],
    images=image_inputs,
    videos=video_inputs,
    padding=True,
    return_tensors="pt",
)
inputs = inputs.to("cuda")

# Inference: Generation of the output
generated_ids = model.generate(**inputs, max_new_tokens=128)
generated_ids_trimmed = [
    out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
]
output_text = processor.batch_decode(
    generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
)
print(output_text)

---

Training Details

Setting	Value
LoRA Rank	16
LoRA Alpha	32
Learning Rate	2e-4
Batch Size	2 (with grad accumulation 8)
Precision	4-bit NF4
Epochs	1
VRAM Used	~5.7GB
Training Duration	49 mins (local)
Gradient Checkpointing	Enabled

Evaluation Metrics

Metric	Value
Token Accuracy	~99.8%
Exact Match	~90%
Val Loss	~0.006

---

Requirements

Install the dependencies using:

pip install -r requirements.txt

License

This project is licensed under the MIT License.

Author

Aaghaaz Khan

• Hugging Face: https://huggingface.co/aaghaazkhan
• LinkedIn: https://www.linkedin.com/in/aaghaaz-khan-778b372a8