Probing Without Labels
Contrast-only Reproductions of Interpretability Results
Negation and truth
Four statements per city: affirmed or negated, with the correct or an incorrect country. On the first three components the four statement types form a tetrahedron. The animation looks down each contrast axis in turn, where that contrast collapses and the other two form a square, then turns the solid freely. Below, drag to rotate; the buttons look down the axis that collapses one contrast.
Llama-2-7B · g-tcPCA · animated
Llama-2-7B · g-tcPCA
Llama-2-7B · tcPCA
Llama-3.1-8B · g-tcPCA
Llama-3.1-8B · tcPCA
True, false, neither
Statements about real cities (true or false) and invented ones (neither), on the first two components.
Llama-2-7B · g-tcPCA
Llama-2-7B · tcPCA
Llama-3.1-8B · g-tcPCA
Llama-3.1-8B · tcPCA
Days of the week
Tuples that differ only in the weekday; the class means are joined in calendar order.
Llama-2-7B · g-tcPCA
Llama-2-7B · tcPCA
Llama-3.1-8B · g-tcPCA
Llama-3.1-8B · tcPCA
Months of the year
Tuples that differ only in the month.
Llama-2-7B · g-tcPCA
Llama-2-7B · tcPCA
Llama-3.1-8B · g-tcPCA
Llama-3.1-8B · tcPCA
Choice of layer
Accuracy against layer for tcPCA, g-tcPCA, CCS and a supervised logistic probe, per dataset and token (choose them above the plot). The paper summarises these as an average over datasets; block 15 (the paper's layer 16) sits on the plateau.
Llama-2-7B
Llama-3.1-8B
Eigenvalue spectra
The top ten tcPCA eigenvalues per dataset and token. One eigenvalue standing out means the pairs isolate a single contrastive direction.
Llama-2-7B
Llama-3.1-8B
CCS accuracy per seed
Every one of the 30 CCS seeds, per dataset and token. Where the seeds split into two groups, a mean describes no run that actually happened.