Outputs
repralign writes several output types during an experiment.
Feature Caches
Extension:
.npz
These contain pooled feature matrices keyed by layer name. They are intended to avoid repeated forward passes when you want to recompute metrics or regenerate plots.
Metadata
Extension:
.json
Metadata files store information such as:
- model name
- adapter name
- selected layers
- pooling mode
- normalization choice
- dataset settings when present
Similarity Tables
Extension:
.csv
These tables are the main intermediate artifact for layer-wise analysis. Each row typically contains:
- metric name
- candidate layer index
- candidate layer name
- reference layer name
- score
Plots
Extension:
.png
The plotting command renders one or more similarity curves into a publication-style figure.