Decode layer activations into per-layer vocabulary predictions for Transformer interpretability.
npx skills add https://github.com/ndif-team/skills --skill logit-lens
Please help me install this Agent Skill. Skill: logit-lens Source: https://github.com/ndif-team/skills/tree/main/plugins/nnsight/skills/logit-lens Command: npx skills add https://github.com/ndif-team/skills --skill logit-lens
Decodes intermediate layer activations into vocabulary predictions to reveal what the model "thinks" at each processing step, offering insight beyond the final output.
Prompt the model with a sample input and run a layer-wise trace to display the top token predictions at each layer.
High-intent search queries and answers about installing and using this skill.
To visualize layer-wise predictions, decode intermediate transformer activations into per-layer vocabulary tokens by applying final normalization and unembedding, revealing what the model computes at each processing step.
The logit lens technique for model interpretability decodes hidden states at each layer to show evolving token predictions, offering insight into how a transformer model computes output beyond the final layer.
Debug information flow across transformer layers by decoding and comparing per-layer vocabulary predictions to diagnose how specific tokens evolve and recall dynamics shift during processing.
Yes, decoding per-layer predictions requires access to intermediate hidden states, the final projection head, and a tokenizer to map intermediate activations back to readable vocabulary words.
Yes, you can use layer-wise decoding for model auditing and interpretability research by tracing how top token predictions evolve across varying prompts to reveal internal recall dynamics.
The best way to trace evolving predictions is applying final normalization and unembedding to hidden states at each layer, which maps intermediate activations to top vocabulary tokens for comparison.
ndif-teamPatch transformer activations across layers, heads, and positions to identify causal components.
one-ieVisualize envelope chains as interactive node graphs with ReactFlow.
tomlupoExplain model predictions by attributing outputs to input features with SHAP values.
dawsonblockTrain sparse autoencoders on transformer activations with SAELens.
