Qwen2.5-VL-7B-TableRecog

Qwen2.5-VL-7B-TableRecog

Model Description

This model is a fine-tuned version of Qwen/Qwen2.5-VL-7B-Instruct specialized for table recognition tasks. The model was trained using LoRA (Low-Rank Adaptation) and then merged with the base model for efficient inference.

Key Features:

• 🔍 Optimized for table structure recognition and content extraction
• 📊 Supports complex table layouts with merged cells and nested structures
• 🌏 Bilingual support (Korean & English)
• ⚡ Full model merge (no adapter required for inference)

Training Details

Base Model

• Model: Qwen2.5-VL-7B-Instruct (7B parameters)
• Architecture: Vision-Language Model with qwen2_5_vl architecture
• Precision: bfloat16

Fine-tuning Configuration

• Method: LoRA (Low-Rank Adaptation)
• Checkpoint: checkpoint-23900 (lora_1010_v64)
• Training Focus: Table recognition and structure understanding
• Dataset: Custom Korean/English table recognition dataset

Merge Details

• Merge Method: PEFT merge_and_unload()
• Output Format: Safetensors (4 shards, ~16GB total)
• Device: Auto device mapping for optimal performance

Usage

import torch
from transformers import AutoModelForVision2Seq, AutoProcessor
from PIL import Image

# Load model and processor
model = AutoModelForVision2Seq.from_pretrained(
    "korea-deep-learning/Qwen2.5-VL-7B-TableRecog",
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True
)

processor = AutoProcessor.from_pretrained(
    "korea-deep-learning/Qwen2.5-VL-7B-TableRecog",
    trust_remote_code=True
)

# Prepare image and text
image = Image.open("table_image.png")
text = "Extract the table content and structure from this image."

# Process inputs
inputs = processor(text=text, images=image, return_tensors="pt").to(model.device)

# Generate output
with torch.no_grad():
    outputs = model.generate(**inputs, max_new_tokens=1024)

result = processor.decode(outputs[0], skip_special_tokens=True)
print(result)

Performance

The model has been evaluated on table recognition tasks with the following metrics:

• OCR Accuracy: 96.96%
• Table Structure Recognition (TSR): 92.79%
• Layout Detection: 100%
• Image Detection: 100%

Use Cases

• 📄 Document digitization with table extraction
• 🏢 Invoice and receipt processing
• 📊 Financial report analysis
• 🗂️ Form data extraction
• 📋 Medical record table recognition

Limitations

• Optimized primarily for Korean and English text
• Best performance on structured tables with clear boundaries
• May require additional fine-tuning for domain-specific terminology
• Large model size (~16GB) requires adequate GPU memory for inference

Training Environment

• Infrastructure: HPC GPU cluster (hpc-gpu1)
• Container: LLaMAFactory Docker environment
• Framework: Transformers 4.49.0, PEFT 0.13.2
• Python Version: 3.11

Citation

@misc{qwen2.5-vl-tablerecog,
  author = {Korea Deep Learning Team},
  title = {Qwen2.5-VL-7B-TableRecog: Table Recognition Model},
  year = {2025},
  publisher = {HuggingFace},
  howpublished = {\url{https://huggingface.co/korea-deep-learning/Qwen2.5-VL-7B-TableRecog}},
}

Acknowledgements

• Base model: Qwen/Qwen2.5-VL-7B-Instruct by Alibaba Cloud
• Training support: 고성능컴퓨팅 지원사업 (HPC Support Project)

License

This model inherits the Apache 2.0 license from the base Qwen2.5-VL-7B-Instruct model.