eval/transfer.md

base rows 1264, new rows 1264, paired 1154

slice	n	base acc	new acc	Δacc [95% CI]	base NLL	new NLL	ΔNLL [95% CI]	new-only/base-only correct
variant=clean	1046	78.68	78.39	-0.29 [-1.75, +1.05]	0.587	0.664	+0.077 [+0.048, +0.109]	22/25
task=buried_emotion	20	75.00	70.00	-5.00 [-15.00, +0.00]	0.988	1.376	+0.387 [-0.063, +0.973]	0/1
task=buried_paws	20	70.00	70.00	+0.00 [+0.00, +0.00]	0.985	1.144	+0.158 [-0.027, +0.405]	0/0
task=buried_qnli	20	90.00	95.00	+5.00 [+0.00, +15.00]	0.256	0.203	-0.053 [-0.140, +0.042]	1/0
task=buried_tweet_offensive	20	75.00	80.00	+5.00 [+0.00, +15.00]	0.430	0.391	-0.038 [-0.077, -0.004]	1/0
task=composition_held_and_or	32	100.00	90.62	-9.38 [-28.12, +0.00]	0.125	0.154	+0.029 [-0.026, +0.120]	0/3
task=composition_held_conditional	32	68.75	68.75	+0.00 [-18.75, +18.75]	0.582	0.556	-0.026 [-0.117, +0.049]	3/3
task=composition_held_or_not	32	90.62	96.88	+6.25 [+0.00, +15.62]	0.370	0.314	-0.056 [-0.133, +0.015]	2/0
task=contrastive_authorization	40	100.00	100.00	+0.00 [+0.00, +0.00]	0.015	0.001	-0.014 [-0.022, -0.007]	0/0
task=contrastive_deadline	40	52.50	52.50	+0.00 [+0.00, +0.00]	1.059	1.827	+0.769 [+0.450, +1.112]	0/0
task=emotion	80	86.25	85.00	-1.25 [-7.50, +3.75]	0.330	0.297	-0.033 [-0.090, +0.025]	2/3
task=mmlu	80	77.50	75.00	-2.50 [-6.25, +0.00]	0.609	0.697	+0.088 [+0.012, +0.176]	0/2
task=mmlu_pro	200	58.50	56.50	-2.00 [-5.00, +0.50]	1.300	1.406	+0.106 [+0.028, +0.197]	2/6
task=paws	80	75.00	78.75	+3.75 [+0.00, +8.75]	0.712	0.758	+0.047 [-0.028, +0.135]	3/0
task=qnli	80	95.00	90.00	-5.00 [-10.00, -1.25]	0.209	0.267	+0.058 [-0.004, +0.129]	0/4
task=sciq	80	97.50	97.50	+0.00 [+0.00, +0.00]	0.048	0.043	-0.004 [-0.014, +0.007]	0/0
task=tweet_offensive	80	81.25	81.25	+0.00 [-3.75, +3.75]	0.354	0.353	-0.001 [-0.018, +0.017]	1/1