opendatalab/TRivia-3B
TRivia-3B is a 3 billion parameter vision-language model developed by opendatalab, fine-tuned from Qwen2.5-VL-3B using the TRivia self-supervised framework for advanced table recognition. It excels at extracting structured data from diverse table formats, including digital, scanned, and photographed tables, with a context length of 32768 tokens. This model is specifically optimized for robust table recognition across various real-world benchmarks, demonstrating superior performance compared to other general-purpose and document-parsing VLMs.
Loading preview...
TRivia-3B: Advanced Table Recognition VLM
TRivia-3B is a 3 billion parameter vision-language model (VLM) developed by opendatalab, specifically designed for robust table recognition. It is fine-tuned from Qwen2.5-VL-3B using the novel TRivia self-supervised framework, which leverages unlabeled table images to enhance its capabilities. This model demonstrates superior performance across multiple real-world table recognition benchmarks, including OmniDocBench v1.5, CC-OCR, and OCRBench v2.
Key Capabilities
- Powerful Table Recognition: Generalizes effectively across digital, scanned, and photographed tables.
- Self-supervised Fine-tuning: Utilizes a reproducible training framework that pushes the boundaries of table recognition using unlabeled data.
- High Performance: Achieves an overall TEDS score of 89.88 and S-TEDS score of 93.60 across key benchmarks, outperforming many expert TR models and larger general-purpose VLMs.
- Structured Output: Supports table parsing with image input and outputs OTSL (Open Table Structure Language) tags, which can be converted to HTML.
Good For
- Document Processing: Extracting structured data from various document types containing tables.
- Data Digitization: Automating the conversion of visual table data into machine-readable formats.
- Research and Development: Exploring advanced self-supervised fine-tuning techniques for vision-language models in specialized domains.
Note: TRivia-3B is an experimental model and currently does not support parsing formulas in tables or tables with embedded images.