spb/wpai0001_uqo

1% source: experiments/summaries/tier1_taskvector_eb5d091bf5ed/results.json2\begin{tabular}{llcccccc}3\toprule4Model & task & zero-shot & few-shot & best layer & vector added (zero-shot) & random vector & few-shot minus direction \\5\midrule6llama3b (28 layers, $d=3072$) & antonym & 0.19 & 0.88 & L14 & 0.84 & 0.06 & 0.91 \\7llama3b (28 layers, $d=3072$) & capital & 0.00 & 1.00 & L14 & 1.00 & 0.00 & 1.00 \\8llama3b (28 layers, $d=3072$) & morph & 0.00 & 0.38 & L2 & 0.00 & 0.00 & 0.33 \\9llama3b (28 layers, $d=3072$) & addthree & 0.00 & 1.00 & L14 & 0.77 & 0.00 & 1.00 \\10qwen3b (36 layers, $d=2048$) & antonym & 0.28 & 0.94 & L25 & 0.88 & 0.06 & 0.97 \\11qwen3b (36 layers, $d=2048$) & capital & 0.00 & 1.00 & L25 & 0.82 & 0.00 & 1.00 \\12qwen3b (36 layers, $d=2048$) & morph & 0.00 & 0.40 & L2 & 0.00 & 0.00 & 0.40 \\13qwen3b (36 layers, $d=2048$) & addthree & 0.00 & 1.00 & L2 & 0.00 & 0.00 & 1.00 \\14qwen7b (28 layers, $d=3584$) & antonym & 0.19 & 0.97 & L16 & 0.91 & 0.16 & 0.97 \\15qwen7b (28 layers, $d=3584$) & capital & 0.23 & 1.00 & L11 & 1.00 & 0.00 & 1.00 \\16qwen7b (28 layers, $d=3584$) & morph & 0.00 & 0.53 & L14 & 0.03 & 0.00 & 0.53 \\17qwen7b (28 layers, $d=3584$) & addthree & 0.00 & 1.00 & L2 & 0.00 & 0.00 & 1.00 \\18\bottomrule19\end{tabular}20