| --- |
| title: Debug-XAI |
| emoji: 🔍 |
| colorFrom: indigo |
| colorTo: blue |
| sdk: docker |
| pinned: false |
| license: apache-2.0 |
| short_description: An Interactive Framework for LLM Failure Analysis |
| suggested_hardware: t4-small |
| --- |
| |
| # Debug-XAI |
|
|
| An Interactive Framework for LLM Failure Analysis using Attribution Methods (Attn-LRP, CP-LRP, Gradient). |
|
|
| ## Features |
|
|
| - **Model Loading**: Load HuggingFace models with optional 4-bit quantization |
| - **Error Token Localization**: Identify erroneous tokens via LLM validators or manual truncation |
| - **Input Attribution**: Visualize token-level attribution using Attn-LRP, CP-LRP, or vanilla gradients |
| - **Perturbation Evaluation**: Evaluate attribution quality by zeroing out top-attributed tokens |
| - **Circuit Visualization**: Build and visualize attribution graphs across model layers |
|
|
| ## Environment Variables (Optional) |
|
|
| - `OPENAI_API_KEY`: OpenAI API key for LLM-based error token localization |
| - `OPENAI_BASE_URL`: Custom OpenAI-compatible API base URL |
| - `LLM_MODELS`: Comma-separated list of model names for validators (default: `gpt-4o-mini`) |
|
|
| Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference |
