|
--- |
|
license: apache-2.0 |
|
language: |
|
- en |
|
base_model: |
|
- Qwen/Qwen2-VL-2B-Instruct |
|
pipeline_tag: image-text-to-text |
|
library_name: transformers |
|
tags: |
|
- latex |
|
- vLM |
|
- Vision |
|
- Codec |
|
--- |
|
|
|
 |
|
|
|
-------------- |
|
|
|
# **LatexMind-2B-Codec** |
|
|
|
The **LatexMind-2B-Codec** model is a fine-tuned version of Qwen2-VL-2B-Instruct, optimized for Optical Character Recognition (OCR), **image-to-text conversion**, and **mathematical expression extraction with LaTeX formatting**. This model integrates a conversational approach with visual and textual understanding to handle multi-modal tasks effectively. |
|
|
|
# Key Enhancements: |
|
|
|
* **SoTA understanding of images with various resolutions & aspect ratios**: LatexMind-2B-Codec achieves state-of-the-art performance on visual understanding benchmarks, including MathVista, DocVQA, RealWorldQA, MTVQA, etc. |
|
|
|
* **Advanced LaTeX extraction**: The model specializes in extracting structured mathematical expressions from images and documents, converting them into LaTeX format for precise rendering and further computation. |
|
|
|
* **Understanding long-duration videos (20min+)**: LatexMind-2B-Codec can process videos over 20 minutes long, enabling high-quality video-based question answering, mathematical solution explanation, and educational content creation. |
|
|
|
* **Agent capabilities for automated operations**: With complex reasoning and decision-making abilities, the model can be integrated with mobile devices, robots, and assistive technologies to automate tasks based on visual and textual inputs. |
|
|
|
* **Multilingual Support**: To serve global users, in addition to English and Chinese, the model supports text recognition inside images across multiple languages, including European languages, Japanese, Korean, Arabic, Vietnamese, etc. |
|
|
|
This model is particularly effective in **retrieving mathematical notations and equations** from scanned documents, whiteboard images, and handwritten notes, ensuring accurate conversion to LaTeX code for further academic and computational applications. |
|
|
|
# Sample Inference with Doc |
|
|
|
 |
|
|
|
Demo: https://huggingface.co/prithivMLmods/LatexMind-2B-Codec/blob/main/latexmind/latexmind-codec.ipynb |
|
|
|
# Use it with Transformers |
|
|
|
|
|
```python |
|
from transformers import Qwen2VLForConditionalGeneration, AutoTokenizer, AutoProcessor |
|
from qwen_vl_utils import process_vision_info |
|
|
|
# default: Load the model on the available device(s) |
|
model = Qwen2VLForConditionalGeneration.from_pretrained( |
|
"prithivMLmods/LatexMind-2B-Codec", torch_dtype="auto", device_map="auto" |
|
) |
|
|
|
# We recommend enabling flash_attention_2 for better acceleration and memory saving, especially in multi-image and video scenarios. |
|
# model = Qwen2VLForConditionalGeneration.from_pretrained( |
|
# "prithivMLmods/LatexMind-2B-Codec", |
|
# torch_dtype=torch.bfloat16, |
|
# attn_implementation="flash_attention_2", |
|
# device_map="auto", |
|
# ) |
|
|
|
# default processer |
|
processor = AutoProcessor.from_pretrained("prithivMLmods/Qwen2-VL-OCR-2B-Instruct") |
|
|
|
# The default range for the number of visual tokens per image in the model is 4-16384. You can set min_pixels and max_pixels according to your needs, such as a token count range of 256-1280, to balance speed and memory usage. |
|
# min_pixels = 256*28*28 |
|
# max_pixels = 1280*28*28 |
|
# processor = AutoProcessor.from_pretrained("Qwen/Qwen2-VL-2B-Instruct", min_pixels=min_pixels, max_pixels=max_pixels) |
|
|
|
messages = [ |
|
{ |
|
"role": "user", |
|
"content": [ |
|
{ |
|
"type": "image", |
|
"image": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-VL/assets/demo.jpeg", |
|
}, |
|
{"type": "text", "text": "Describe this image."}, |
|
], |
|
} |
|
] |
|
|
|
# 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) |
|
``` |
|
# Buf |
|
```python |
|
buffer = "" |
|
for new_text in streamer: |
|
buffer += new_text |
|
# Remove <|im_end|> or similar tokens from the output |
|
buffer = buffer.replace("<|im_end|>", "") |
|
yield buffer |
|
``` |
|
|
|
# Intended Use |
|
|
|
**LatexMind-2B-Codec** is designed for tasks that require **image-based text recognition**, **math equation extraction**, and **multi-modal understanding**. It is particularly useful in the following scenarios: |
|
|
|
**Optical Character Recognition (OCR)** – Extracting printed and handwritten text from images, documents, and scanned pages. |
|
**Math Expression Recognition** – Converting mathematical notations into structured **LaTeX format** for further computation and documentation. |
|
**Image-to-Text Conversion** – Generating accurate descriptions for text-rich and math-heavy images. |
|
**Document and Academic Processing** – Assisting researchers, students, and professionals in digitizing handwritten notes and extracting structured content from books, PDFs, and whiteboards. |
|
**Automated Educational Support** – Enabling AI-powered tutors, content summarization, and interactive learning for subjects involving complex equations. |
|
**Multi-Language OCR** – Recognizing text inside images across multiple languages, including English, Chinese, Japanese, Korean, Arabic, and various European languages. |
|
**Video-Based Question Answering** – Understanding long-duration videos for content summarization, question answering, and structured data extraction. |
|
|
|
# Limitations |
|
|
|
Despite its capabilities, **LatexMind-2B-Codec** has some inherent limitations: |
|
|
|
**Handwritten Text Accuracy** – While it can recognize handwritten equations, performance may degrade with highly unstructured or messy handwriting. |
|
**Complex LaTeX Formatting** – The model may struggle with deeply nested or ambiguous LaTeX expressions, requiring manual corrections for precise formatting. |
|
**Low-Resolution Images** – Extracting accurate text from blurry or low-resolution images can lead to misinterpretations or OCR errors. |
|
**Contextual Understanding in Multi-Step Equations** – While it recognizes math expressions, solving multi-step problems autonomously may be limited. |
|
**Limited Support for Rare Mathematical Notations** – Some specialized or domain-specific symbols may not be recognized with high accuracy. |
|
**Processing Speed for Large Documents** – Performance may slow down when handling extremely large documents or dense mathematical content in real-time applications. |
|
**Language-Specific OCR Variability** – While it supports multiple languages, OCR accuracy may vary depending on the script complexity and font style. |
