| % Main results: raw conditional-loglikelihood accuracy (acc), 0-shot, bf16, Ascend 910C. | |
| % Random baseline = 25.0 (all tasks are 4-option multiple choice). Belebele diff = zh - en, percentage points. | |
| \begin{table}[t] | |
| \centering | |
| \caption{Zero-shot multiple-choice accuracy (\%) of base models under conditional log-likelihood scoring. C3 is Chinese reading comprehension (dialogue / mixed / question-count-weighted total); Belebele is a parallel passage-based RC benchmark (900 items per language, identical passages and questions across languages); CMMLU is a 67-subject Chinese knowledge benchmark (macro-average over subjects). All values are the harness-native \texttt{acc} (unnormalized log-likelihood argmax); length-normalized \texttt{acc\_norm} is reported in the appendix. Chance level is 25\%.} | |
| \label{tab:main} | |
| \small | |
| \begin{tabular}{lcccccccc} | |
| \toprule | |
| & & \multicolumn{3}{c}{C3 (zh)} & \multicolumn{2}{c}{Belebele} & Diff & CMMLU \\ | |
| \cmidrule(lr){3-5}\cmidrule(lr){6-7}\cmidrule(lr){8-8} | |
| Model & Params & dial. & mixed & total & zh & en & zh$-$en & macro \\ | |
| & & \multicolumn{7}{c}{accuracy (\%)} \\ | |
| \midrule | |
| Haidass1.5-143M & 0.14B & 42.91 & 41.06 & 41.96 & 22.89 & 24.44 & -1.56 & 25.44 \\ | |
| SmolLM2-135M & 0.13B & 36.46 & 39.11 & 37.82 & 22.67 & 24.22 & -1.56 & 24.89 \\ | |
| Qwen2.5-0.5B & 0.49B & 48.57 & 49.35 & 48.97 & 44.44 & 45.78 & -1.33 & 48.80 \\ | |
| Qwen3-0.6B-Base & 0.60B & 52.54 & 52.30 & 52.42 & 60.67 & 56.56 & 4.11 & 51.18 \\ | |
| \midrule | |
| GPT-X2.5-135M & 0.14B & 37.14 & 39.01 & 38.10 & 22.89 & 26.44 & -3.56 & 25.27 \\ | |
| cagliostro-v3 & 0.15B & 36.61 & 38.66 & 37.67 & 26.00 & 27.11 & -1.11 & 25.26 \\ | |
| MobileLLM-R1-140M-base & 0.14B & 33.60 & 34.17 & 33.89 & 24.33 & 25.22 & -0.89 & 25.28 \\ | |
| \bottomrule | |
| \end{tabular} | |
| \end{table} | |