SPB Git

spb/wp31_uqo

Public
2.4 KB · 49 lines latex
1\begin{table}[t]\centering\small2\caption{103M pilot: held-out cross-entropy (nats/token) and perplexity on 327,680 FineWeb-Edu validation tokens and3on the LOCKED test split (40 $\times$ 16 $\times$ 512 tokens each) under forced random budgets. Budget 0 of the adapters4equals the backbone bitwise.}5\label{tab:pilotppl}6\begin{tabular}{lllrr}\toprule7System & split & policy & loss & ppl\\\midrule8backbone 103m & val & forced b0 & 3.8746 & 48.16\\9backbone 103m & val & forced b1 & 3.8746 & 48.16\\10backbone 103m & val & forced b2 & 3.8746 & 48.16\\11backbone 103m & val & forced b4 & 3.8746 & 48.16\\12backbone 103m & val & forced b8 & 3.8746 & 48.16\\13enlarged 102m & val & forced b0 & 3.8689 & 47.89\\14enlarged 102m & val & forced b1 & 3.8689 & 47.89\\15enlarged 102m & val & forced b2 & 3.8689 & 47.89\\16enlarged 102m & val & forced b4 & 3.8689 & 47.89\\17enlarged 102m & val & forced b8 & 3.8689 & 47.89\\18gr 103m & val & forced b0 & 3.8746 & 48.16\\19gr 103m & val & forced b1 & 3.8731 & 48.09\\20gr 103m & val & forced b2 & 3.8722 & 48.05\\21gr 103m & val & forced b4 & 3.8714 & 48.01\\22gr 103m & val & forced b8 & 3.8718 & 48.03\\23prisme 103m editor & val & forced b0 & 3.8746 & 48.16\\24prisme 103m editor & val & forced b1 & 3.8731 & 48.09\\25prisme 103m editor & val & forced b2 & 3.8722 & 48.05\\26prisme 103m editor & val & forced b4 & 3.8714 & 48.01\\27prisme 103m editor & val & forced b8 & 3.8718 & 48.03\\28backbone 103m & test & forced b0 & 3.8226 & 45.72\\29backbone 103m & test & forced b1 & 3.8226 & 45.72\\30backbone 103m & test & forced b2 & 3.8226 & 45.72\\31backbone 103m & test & forced b4 & 3.8226 & 45.72\\32backbone 103m & test & forced b8 & 3.8226 & 45.72\\33gr 103m & test & forced b0 & 3.8226 & 45.72\\34gr 103m & test & forced b1 & 3.8211 & 45.66\\35gr 103m & test & forced b2 & 3.8203 & 45.62\\36gr 103m & test & forced b4 & 3.8195 & 45.58\\37gr 103m & test & forced b8 & 3.8201 & 45.61\\38prisme 103m editor & test & forced b0 & 3.8226 & 45.72\\39prisme 103m editor & test & forced b1 & 3.8211 & 45.66\\40prisme 103m editor & test & forced b2 & 3.8202 & 45.61\\41prisme 103m editor & test & forced b4 & 3.8195 & 45.58\\42prisme 103m editor & test & forced b8 & 3.8200 & 45.60\\43enlarged 102m & test & forced b0 & 3.8207 & 45.64\\44enlarged 102m & test & forced b1 & 3.8207 & 45.64\\45enlarged 102m & test & forced b2 & 3.8207 & 45.64\\46enlarged 102m & test & forced b4 & 3.8207 & 45.64\\47enlarged 102m & test & forced b8 & 3.8207 & 45.64\\48\bottomrule\end{tabular}\end{table}49
Files · main
Menu

Theme