openai/circuit_sparsity

Open-source release accompanying Gao et al. 2025

What it solves

This project provides tools to inspect and visualize sparse circuit models, helping researchers understand how neural networks perform specific tasks by identifying the minimal set of components (circuits) required for those tasks.

How it works

The project implements a lightweight GPT architecture for inference and a Streamlit-based dashboard. The visualizer loads circuit masks, activations, and importance data from a remote blob store, allowing users to interactively explore how pruning sweeps and node budgets affect model behavior.

Who it’s for

Researchers and developers interested in mechanistic interpretability and the study of sparse circuits in neural networks.

Highlights

  • Interactive Streamlit dashboard with Plotly-rendered plots for circuit exploration.
  • Lightweight GPT implementation optimized for CPU/GPU inference.
  • Access to a suite of released models (including csp_yolo and csp_bridge variants) and pruning sweep data.
  • Built-in utilities for recording activations via hooks and managing cached data artifacts.

Related

  • Project
    micytao/vllm-playground

    A modern web interface for managing and interacting with vLLM servers, supporting multi-instance backends, multimodal generation, and agentic tool integration.

  • Project
    google/orbax

    A checkpointing and persistence library for JAX models that enables efficient saving and restoring of model states, including support for asynchronous checkpointing.

  • Project
    DavidVivancos/Neuraxon

    A bio-inspired computational framework using trinary logic and continuous-time simulation to enable real-time learning and structural plasticity.

  • Project
    ImKKingshuk/LockKnife

    LockKnife is a Python‑orchestrated, Rust‑accelerated framework that unifies Android forensics, credential recovery, APK analysis, runtime instrumentation and exploitation into a single terminal‑based workspace (TUI) with a headless CLI alternative. It supports device extraction via ADB, offline PIN/bruteforce attacks, artifact correlation, report generation, optional Frida‑based runtime hooks, Wi‑Fi/Bluetooth exploitation, and optional AI/threat‑intel modules. Installation is a one‑line script or Homebrew tap; the tool is GPL‑3.0‑only and intended for lawful security research.