File size: 1,816 Bytes
5463614 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 | \begin{table}[t]
\centering
\small
\begin{tabular}{lllrp{0.38\linewidth}}
\toprule
Audit & Kind & Suite & Changed & Metrics \\
\midrule
real-instruct-qwen2.5-0.5b & instruction tuning & lite & 6 & ifeval\_prompt\_accuracy, xstest\_safe\_over\_refusal\_rate, or\_bench\_hard\_over\_refusal\_rate, xstest\_unsafe\_compliance\_rate, or\_bench\_toxic\_compliance\_rate, output\_divergence\_kl \\
null-identity-s0 & null identity & lite & 0 & none \\
null-identity-s1 & null identity & lite & 0 & none \\
null-micro-noise-s0 & null micro noise & lite & 1 & output\_divergence\_kl \\
null-micro-noise-s1 & null micro noise & lite & 1 & output\_divergence\_kl \\
real-toolsure-orpo & orpo & lite & 1 & output\_divergence\_kl \\
planted-noise-0.10 & planted noise & lite & 3 & xstest\_unsafe\_compliance\_rate, or\_bench\_toxic\_compliance\_rate, output\_divergence\_kl \\
planted-noise-0.40 & planted noise & lite & 9 & hellaswag\_ece, gsm8k\_accuracy, mgsm\_accuracy, ifeval\_prompt\_accuracy, xstest\_safe\_over\_refusal\_rate, or\_bench\_hard\_over\_refusal\_rate, xstest\_unsafe\_compliance\_rate, or\_bench\_toxic\_compliance\_rate, output\_divergence\_kl \\
planted-prune-0.30 & planted prune & lite & 3 & gsm8k\_accuracy, or\_bench\_hard\_over\_refusal\_rate, output\_divergence\_kl \\
planted-zero-last-mlp & planted zero & lite & 1 & output\_divergence\_kl \\
real-shortthink-sft & sft & lite & 3 & arc\_challenge\_ece, ifeval\_prompt\_accuracy, output\_divergence\_kl \\
real-bitnarrow-qwen7b & weight edit & lite & 2 & xstest\_unsafe\_compliance\_rate, output\_divergence\_kl \\
\bottomrule
\end{tabular}
\caption{Audits in the study. A metric counts as changed when its paired interval excludes zero after Benjamini Hochberg correction and the change exceeds its minimum effect.}
\label{tab:audits}
\end{table}
|