Inspect and modify neural network activations during forward passes.
npx skills add https://github.com/ndif-team/skills --skill nnsight-basics
Please help me install this Agent Skill. Skill: nnsight-basics Source: https://github.com/ndif-team/skills/tree/main/plugins/nnsight/skills/nnsight-basics Command: npx skills add https://github.com/ndif-team/skills --skill nnsight-basics
NNsight Basics enables researchers and engineers to access and manipulate neural network internals during forward passes, accelerating debugging, hypothesis testing, and targeted interventions.
Install the nnsight package and start a tracing session to inspect layer outputs and saved tensors.
High-intent search queries and answers about installing and using this skill.
To access and save intermediate activations during a forward pass, you can start a tracing session to inspect layer outputs and save tensors. This allows you to capture internal model states for analysis.
Yes, you can intervene on neural network activations during tracing to run controlled experiments. You can zero out, add noise to, or replace values in the activations to test specific hypotheses.
Gradient access and generation workflows are supported for large language models. You can trace information flow and compute gradients during forward passes to study model behavior.
The best way to trace information flow across multiple prompts is using multi-prompt tracing. This technique allows you to study how information moves through the model by comparing activation patterns across different inputs.
Yes, you can use this for model debugging and feature attribution checks on vision transformers. It enables you to inspect module outputs and intervene on activations to identify and fix model issues.
galeepLoad pre-trained Transformer models from the Hugging Face Hub for NLP, vision, audio, and multimodal tasks.
qcmuuInspect and causally manipulate GPT-style transformer activations with TransformerLens.
hhhi21gPerform causal interventions on PyTorch model activations.
pacphiDefine and train PyTorch neural networks with CUDA acceleration.
