A 49M-parameter Kimi K3-inspired language model trainable on one 8GB GPU

Language model tokenization at GB/s

A Claude Code skill that encodes battle-tested editorial principles, section-specific rhetorical moves, and a structured writing pipeline for research papers. Brainstorm → Draft 0 → Evaluate → Writ…

Companion code for the global workspace interpretability paper

Reinforcement Learning from Text Feedback

A library for generative social simulation

A fast, helpful, and open-source document parser

Code for Negation Neglect

Official code for the paper: "Simple LLM Baselines are Competitive for Model Diffing"

Training LLMs to Report Their Learned Behaviors

Extract residual-stream activations and apply steering vectors (including activation oracles) to any vLLM model during inference.

A toolkit that provides a range of model diffing techniques including a UI to visualize them interactively.

A markdown native slides tool for academics building with agents.

Neural building blocks for speaker diarization: speech activity detection, speaker change detection, overlapped speech detection, speaker embedding

Reproducing "Natural Emergent Misalignment from Reward Hacking" (MacDiarmid et al., Anthropic 2025) with open-source models. Includes reward-hackable RL environments, misalignment evaluations, trai…

The nnsight package enables interpreting and manipulating the internals of deep learned models.

Research project seeking to confirm/disprove the manifold hypothesis in the context of LLM internal activations

Interpretability Research Tool Kit

A simplified and more accurate next token Patchscope

A carefully curated collection of high-quality libraries, projects, tutorials, research papers, and other essential resources focused on Mechanistic Interpretability, a growing subfield in machine …

Code for `LLM2VEC-GEN: Generative Embeddings from Large Language Models`

Late Interaction Models Training & Retrieval

Optimize prompts, code, and more with AI-powered Reflective Optimization

Official PyTorch implementation for "TensorLens: End-to-End Transformer Analysis via High-Order Attention Tensors" [ACL 2026]

🪄 Interpreto is an interpretability toolbox for LLMs

ADAG: Transluce's MLP neuron-level circuit tracing library

A collection of lightweight interpretability scripts to understand how LLMs think

Pivotal Token Search