aaghaazkhan/Qwen2_5_3B_VL_HandWritten_LaTeX_OCR

All training and inference code is available in the GitHub repository:

https://github.com/aaghaazkhan/Qwen2_5_3B_VL_HandWritten_LaTeX_OCR

---

Model Description

This model converts images of handwritten mathematical expressions into LaTeX code using a fine-tuned version of Qwen 2.5-VL. It was trained using LoRA on a 4-bit quantized model setup, making it GPU-friendly and accessible for low-resource environments.

Ideal for automating LaTeX conversion in:

1. Math learning apps
2. Academic tools
3. E-notes platforms
4. Personal study or research agents

Hardware Used

Trained locally on NVIDIA RTX 3050 (6GB VRAM)

Total VRAM usage: 5.72 GB

---

Using 🤗 Transformers to Chat

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(
    "aaghaazkhan/Qwen2_5_3B_VL_HandWritten_LaTeX_OCR", 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

Author

Aaghaaz Khan

• GitHub: https://github.com/aaghaazkhan
• LinkedIn: https://www.linkedin.com/in/aaghaaz-khan-778b372a8