docscopeOCR-7B-050425-exp

The docscopeOCR-7B-050425-exp model is a fine-tuned version of Qwen/Qwen2.5-VL-7B-Instruct, optimized for Document-Level Optical Character Recognition (OCR), long-context vision-language understanding, and accurate image-to-text conversion with mathematical LaTeX formatting. Built on top of the Qwen2.5-VL architecture, this model significantly improves document comprehension, structured data extraction, and visual reasoning across diverse input formats.

Key Enhancements

Advanced Document-Level OCR: Capable of extracting structured content from complex, multi-page documents such as invoices, academic papers, forms, and scanned reports.
Enhanced Long-Context Vision-Language Understanding: Designed to handle dense document layouts, long sequences of embedded text, tables, and diagrams with coherent cross-reference understanding.
State-of-the-Art Performance Across Resolutions: Achieves competitive results on OCR and visual QA benchmarks such as DocVQA, MathVista, RealWorldQA, and MTVQA.
Video Understanding up to 20+ minutes: Supports detailed comprehension of long-duration videos for content summarization, Q&A, and multi-modal reasoning.
Visually-Grounded Device Interaction: Enables mobile/robotic device operation via visual inputs and text-based instructions using contextual understanding and decision-making logic.

Quick Start with Transformers

python

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

model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
    "prithivMLmods/docscopeOCR-7B-050425-exp", torch_dtype="auto", device_map="auto"
)

processor = AutoProcessor.from_pretrained("prithivMLmods/docscopeOCR-7B-050425-exp")

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."},
        ],
    }
]

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")

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

Parameter	Value
Dataset Size	274,209 samples (Modular Combination of Datasets)
Model Architecture	`Qwen2_5_VLForConditionalGeneration`
Hardware	2 × NVIDIA A100 SXM (32 vCPUs)
Total Disk	170,000 MB
Training Time	9,020 seconds (~2.51 hours)
Learning Rate	1e-5
Scheduler	Linear Decay
Warmup Steps	750
Precision	bfloat16

[!note] The open dataset image-text response will be updated soon.

Intended Use

This model is intended for:

High-fidelity OCR from documents, forms, receipts, and printed or scanned materials.
Image and document-based question answering for educational and enterprise applications.
Extraction and LaTeX formatting of mathematical expressions from printed or handwritten content.
Retrieval and summarization from long documents, slides, and multi-modal inputs.
Multilingual OCR and structured content extraction for global use cases.
Robotic or mobile automation with vision-guided contextual interaction.

Limitations

May show degraded performance on extremely low-quality or occluded images.
Not optimized for real-time applications on low-resource or edge devices due to computational demands.
Variable accuracy on uncommon or low-resource languages/scripts.
Long video processing may require substantial memory and is not optimized for streaming applications.
Visual token settings affect performance; suboptimal configurations can impact results.
In rare cases, outputs may contain hallucinated or contextually misaligned information.

References

DocVLM: Make Your VLM an Efficient Reader https://arxiv.org/pdf/2412.08746v1
YaRN: Efficient Context Window Extension of Large Language Models
https://arxiv.org/pdf/2309.00071
Qwen2-VL: Enhancing Vision-Language Model’s Perception of the World at Any Resolution
https://arxiv.org/pdf/2409.12191
Qwen-VL: A Versatile Vision-Language Model for Understanding, Localization, Text Reading, and Beyond
https://arxiv.org/pdf/2308.12966
A Comprehensive and Challenging OCR Benchmark for Evaluating Large Multimodal Models in Literacy https://arxiv.org/pdf/2412.02210