openai/transformer-debugger
Transformer Debugger
This project is a language model interpretability tool designed to visualize and analyze the internal activations and attention heads of neural networks. It provides a framework for understanding how models process information by capturing internal model states and rendering them through an interactive web interface.
The system specializes in decomposing high-dimensional activations into interpretable latent features using sparse autoencoders and generating automated natural language explanations for individual model nodes. It enables the discovery of model circuits by mapping connections between components and tracing the pathways responsible for specific behaviors.
The toolset covers a broad range of analysis capabilities, including gradient-based attribution to quantify node importance and real-time interventions such as node ablation to observe how modifying internal states affects model output. It also supports live inference sampling and the retrieval of activation data from remote backend servers and blob storage.
Features
- Language Model Interpretability - Provides a comprehensive framework for explaining the internal processing and outputs of large language models.
- Activation Interventions - Implements real-time modification and ablation of internal neural network activations during the forward pass.
- Automated Neuron Explanations - Generates natural language descriptions for individual model nodes to explain which features or patterns they detect.
- Neuron Inspections - Provides tools for the interactive visual inspection of individual neurons and their behavioral contributions.
- Internal Activation Hooks - Implements hook-based mechanisms to capture internal tensor activations and gradients during the model forward pass.
- Model Activation Viewers - Ships an interactive web interface for displaying real-time activation data from individual model components.
- Sparse Autoencoder Attributors - Uses sparse autoencoders to decompose model activations into interpretable high-level latent features.
- Feature Decomposition - Projects high-dimensional activations into an overcomplete basis to isolate interpretable features from dense layers.
- Model Circuit Discovery - Traces connections between internal components to map the specific pathways and circuits responsible for certain behaviors.
- Activation Hooking - Inserts programmable functions into the execution graph to intercept and capture internal tensors during inference.
- Activation Ablations - Disables specific computation nodes to observe how their absence affects downstream activations and behavior.
- Activation Decompositions - Provides tools to decompose high-dimensional layer activations into interpretable latent features using sparse autoencoders.
- Internal Component Visualizations - Renders a visual interface to inspect individual elements such as neurons, attention heads, and autoencoder latents.
- Activation Tracing - Identifies the set of preceding nodes that caused a specific node to write to the residual stream.
- Node Importance Quantifications - Estimates the effect of a node on a representation direction by calculating the product of its activation and gradient.
- Neuron Activation Visualization - Provides capabilities to identify and collate inputs that maximize specific neuron activations to reveal learned patterns.
- Input Times Gradient Attribution Methods - Quantifies component importance using the product of activations and their corresponding gradients.
- Activation-Capture Sampling - Triggers on-demand text generation and simultaneous activation capture by routing requests to a hosted model.
- Live Activation Fetching - Retrieves real-time activations and loss data for individual neurons by performing live inference on a subject model.
- Interpretability Data Retrieval - Retrieve collated activation datasets for neurons and attention heads to analyze responses to specific tokens.
- Remote Inference Streaming - Streams real-time internal model states from a remote inference backend to a local visualization interface.
Star history
How this analysis was created: This summary and feature list were written by an AI model that read the project's README and public documentation pages. Each feature links to the documentation it came from; stars, license and language come straight from the GitHub API. The model does not read the source code, and the analysis is refreshed when the project is re-analysed. Learn more on our About page.
