NuExtract-2-8B
| Property | Value |
|---|---|
| Model Size | 8B parameters |
| Base Model | InternVL2.5-8B |
| Author | NuMind |
| Model Hub | Hugging Face |
What is NuExtract-2-8B?
NuExtract-2-8B is a specialized model designed for structured information extraction tasks, capable of processing both text and images. Built on the InternVL2.5 architecture, it excels at extracting specific data points according to user-defined JSON templates. The model supports multilingual inputs and can handle complex extraction scenarios with multiple data types.
Implementation Details
The model implements a sophisticated template-based extraction system that supports various data types including verbatim strings, generic strings, integers, numbers, dates, arrays, enums, and multi-label classifications. It uses advanced preprocessing techniques for image handling and supports both single and batch processing of inputs.
- Supports multiple image inputs per prompt with dynamic preprocessing
- Implements flash attention 2 for improved performance
- Handles both zero-shot and few-shot learning scenarios
- Uses bfloat16 precision for efficient computation
Core Capabilities
- Template-based structured information extraction
- Multimodal processing (text and images)
- Multilingual support
- Automatic template generation from various formats
- Batch processing of mixed text and image inputs
- Support for in-context learning with examples
Frequently Asked Questions
Q: What makes this model unique?
NuExtract-2-8B stands out for its specialized focus on structured information extraction using JSON templates, combined with multimodal capabilities. It can process both text and images while maintaining precise control over the output format, making it ideal for automated data extraction systems.
Q: What are the recommended use cases?
The model is particularly well-suited for: document information extraction, image content analysis, form processing, structured data extraction from mixed sources, and automated data entry systems. It's recommended to use the model with temperature set to 0 for optimal extraction accuracy.
