| Property | Value |
|---|---|
| Base Model | vidore/colpali-v1.3-hf |
| Training Framework | PaliGemma-3B |
| Authors | Selim Çavaş & Muhammet Fatih Aktuğ |
| Primary Language | Turkish |
What is turkish-colpali?
turkish-colpali is a specialized fine-tuned version of the ColPali model, specifically designed for Turkish document retrieval. Built upon PaliGemma-3B architecture, it implements an innovative approach that combines both visual and textual features for efficient document indexing and retrieval. The model was trained on carefully curated Turkish textbooks and science magazine content, making it particularly effective for academic and scientific document processing.
Implementation Details
The model employs a sophisticated training strategy utilizing Vision Language Models (VLMs) to generate ColBERT-style multi-vector representations. The training process involved converting PDF documents to page images and using gemini-2.0-flash-exp for synthetic query generation. The implementation supports both textual and visual retrieval capabilities, extending traditional RAG system functionalities.
- Trained with learning rate of 5e-05 and linear scheduler
- Uses ADAMW_TORCH optimizer with specific beta parameters
- Implements batch processing with gradient accumulation
- Supports bfloat16 precision for efficient computation
Core Capabilities
- Dual-modal document indexing (text and visual)
- Efficient retrieval of Turkish academic content
- PDF document processing and analysis
- Multi-vector representation generation
- Support for diverse query types
Frequently Asked Questions
Q: What makes this model unique?
The model's ability to process both visual and textual content in Turkish documents sets it apart, making it especially valuable for comprehensive document analysis. Its fine-tuning on Turkish academic content ensures high performance on educational and scientific materials.
Q: What are the recommended use cases?
The model excels in processing Turkish textbooks, scientific magazines, and well-structured PDF documents. It's particularly suitable for academic content management systems, digital libraries, and educational resource indexing systems where both visual and textual content need to be analyzed.