logit-lens - Decode intermediate layer predictions
Decode intermediate layer activations into vocabulary predictions
Tags
Updated: 2026-03-10Capabilities
Typical Inputs
Typical Outputs
What this skill does
- decode intermediate layer predictions
- apply layer normalization
- apply unembedding head
- extract top predictions
- visualize layer-wise predictions
- compare predictions across layers
- process multiple prompts
Inputs
- language model
- input prompt
- target layers
Outputs
- vocabulary probabilities
- predicted tokens
- layer-wise predictions
- prediction visualization
Requirements
- nnsight library
- PyTorch
- model access