nnsight-remote-interpretability
Category: ml-evaluation-and-safety · License: MIT · Version: 1.0.0
Tags: nnsight, NDIF, Remote Execution, Mechanistic Interpretability, Model Internals
When to use it
Provides guidance for interpreting and manipulating neural network internals using nnsight with optional NDIF remote execution. Use when needing to run interpretability experiments on massive models (70B+) without local GPU resources, or when working with any PyTorch architecture.Example requests
Ask your agent in plain language, for example:- “How do I grab the layer 20 residual stream from Llama-3.1-70B when I don’t have a GPU?”
- “I’m running activation patching on Llama-3.1-8B for my IOI-style prompts in ioi_patch.py. Locally it works, but when I set remote=True my loop over layers keeps timing out, and sometimes I get an error that my saved tensor is a Proxy object outside the with block. How do I fix it?”
- “I want to see what a Mamba state space model represents internally at each block, and copy the hidden state from one prompt into another prompt’s forward pass. I’d rather not rewrite the model in some interpretability library. What approach works?”
Install
/plugin marketplace add KalarisLabs/research-agent-skills and install the whole ml-evaluation-and-safety category with /plugin install ml-evaluation-and-safety@research-agent-skills.
Related
slime-rl-trainingsparse-autoencoder-trainingtransformer-lens-interpretability- Skills for machine learning research