HTML-Pruner-Llama-1B
| Property | Value |
|---|---|
| Parameter Count | 1.24B parameters |
| License | Apache 2.0 |
| Paper | HtmlRAG: HTML is Better Than Plain Text for Modeling Retrieved Knowledge in RAG Systems |
| Base Model | meta-llama/Llama-3.2-1B |
What is HTML-Pruner-Llama-1B?
HTML-Pruner-Llama-1B is a specialized language model designed to enhance RAG (Retrieval-Augmented Generation) systems by optimizing HTML content processing. This 1.24B parameter model implements an innovative two-step HTML pruning approach that maintains semantic integrity while reducing content length for more efficient processing.
Implementation Details
The model employs a sophisticated two-step block-tree-based HTML pruning strategy: first utilizing an embedding model for block scoring, followed by a path generative model for further refinement. It includes a Lossless HTML Cleaning process that preserves semantic information while removing redundant structures.
- Two-Step Block-Tree-Based HTML Pruning architecture
- Lossless HTML Cleaning capability
- Built on LLaMA architecture with BF16 tensor type
- Optimized for context windows up to 60 tokens
Core Capabilities
- Efficient HTML content pruning while maintaining semantic meaning
- Block-tree structure analysis and optimization
- Integration with various embedding models (BM25, BGE, E5-Mistral)
- Competitive performance across multiple QA datasets (ASQA, HotpotQA, NQ, TriviaQA)
Frequently Asked Questions
Q: What makes this model unique?
This model stands out for its specialized HTML processing capabilities, offering a novel two-step pruning approach that outperforms traditional text-based RAG systems. It achieves state-of-the-art results across multiple benchmarks while maintaining HTML structure integrity.
Q: What are the recommended use cases?
The model is ideal for RAG systems requiring HTML document processing, question-answering systems, and applications needing efficient HTML content summarization while preserving semantic structure. It's particularly effective for scenarios where context length optimization is crucial.
