Project DescriptionÂ
Web-based interactive visualization software tool for evaluating and debugging LLM models unique processing of input prompts based on investigating attention weights distribution for tokens of input prompt.
Features
Implemented a token-based attention weight visualization for user-centric understanding of the LLM model’s operations via a dynamic heatmap in React and D3.js, enabling in-depth exploration by toggling across different model layers and attention heads.
Integrated Hugging Face's Transformers library with Flask for real-time extraction and analysis of attention weights.
