Project structure of: openai/automated-interpretability

• automated-interpretability Neuron Explainer AI toolkit, demos & datasets
  • neuron-explainer Neuron Explainer: AI model analysis and demo hub
    • demos Neuron explanation demo hub
      • explain_puzzles.py AI-powered puzzle solver.
      • generate_and_score_explanation.py Neuron data explanation generation and simulation.
      • generate_and_score_token_look_up_table_explanation.py Generate and score token explanation table.
    • neuron_explainer Neuron Explainer: AI model analysis and explanation.
      • activations Neuron activations management and analysis.
        • activation_records.py Activation records management and analysis.
        • activations.py Neuron activation record slicing and processing.
        • token_connections.py Reads token-activation data from Azure, formats results.
      • api_client.py Async OpenAI API client with error handling and exponential backoff
      • azure.py Azure format converter for input URLs.
      • explanations Neuron explanation, AI model, asynchronous loading
        • calibrated_simulator.py Calibrated Neuron Simulator Improvements
        • explainer.py Neuron explainer AI model for generating explanations using API calls and prompts.
        • explanations.py Neuron explanation, asynchronous loading, sorted indices.
        • prompt_builder.py PromptBuilder: Format messages, build prompts with GPT-4 encoding.
        • puzzles.py Puzzle class for ground truth and false explanations.
        • scoring.py Scoring function for neuron evaluations using correlation & R-squared.
        • simulator.py Neuron simulator: improved initialization, API calls, prompt formats.
        • test_explainer.py Test neuron explainer with event loop, GPT-4, and Harmony V4.
        • test_simulator.py Test script for neuron accuracy in neural networks, text-davinci-003 model.
        • token_space_few_shot_examples.py Few-shot medical/scientific token class for neuron explainers.
      • fast_dataclasses Efficient, validated dataclasses with orjson.
        • __init__.py Imports FastDataclass functions and sets all.
        • fast_dataclasses.py Efficient dataclasses with limited validation using orjson.
        • test_fast_dataclasses.py Tests FastDataclass functionality and error handling.
    • README.md Neuron simulation code for explanation generation
    • setup.py Setup script for neuron_explainer package
  • neuron-viewer Neuron viewer website, Flask server, Tailwind CSS, JSON data
    • python
      • server.py Starts Flask server, imports libs, loads JSON data
    • README.md Neuron viewer website with local development support
    • src Visualizes and analyzes neuron data with component-based interface
      • App.jsx Sets up routing with CSS, Feed import.
      • feed.jsx Neuron-Viewer: Display welcome or neuron info based on selection.
      • heatmapGrid.tsx Heatmap grid for TokenAndActivation objects.
      • index.jsx Setup React app with performance measurement
      • interpAPI.ts Retrieves top-connected neurons and layer pairs using JSON loading.
      • panes Visualize and analyze neuron data, component-based interface.
        • datasetList.jsx Visualize and analyze neuron activation datasets.
        • explanation.jsx Neuron viewer: component for loading, toggling sequences
        • index.js Importing components for application use.
        • similarNeurons.jsx Fetch and display neuron info, organized visually with loading animation.
        • topTokens.jsx Interactive token display with red text and tooltips.
      • reportWebVitals.js Measures web vitals with callback function.
      • simulationHeatmap.tsx Visualizes matching sequences in heatmaps, rendering each simulation.
      • tokenHeatmap.tsx Visualize token activations as heatmap with customizable colors and boundaries.
      • types.ts Color interpolation and normalization functions.
      • utils.ts Memoize async, get query params utilities
      • welcome.tsx NeuronForm component, hooks, predefined neurons, random selection.
    • tailwind.config.js Tailwind CSS config with source content and empty extend/plugins.
  • README.md Neuron explanation tools for language models with datasets and fixes.