ScanNet 单图数据(vst_scannet_si_v1)p10 验证 · baseline 对比

同起点(lijia-line SFT ckpt)· 同口径(no-thinking,t0.0,circular)· 1000 步。蓝线 = 纯 baseline;橙线 = +10% 英文版数据;绿线 = +10% 中文版数据(同图同 GT,仅问答语言不同)。step100 属 warmup 段。

感知

立体几何(SolidGeo 943 题)

几何 geo-mc(老题库,你 6 月同款)

全部数值

step感知NP立体几何SolidGeo几何geo-mc
baseline+10%英文+10%中文baseline+10%英文+10%中文baseline+10%英文+10%中文
10021.84%22.57%22.47%25.24%23.86%24.28%26.05%27.95%26.76%
20023.30%22.78%24.03%28.74%30.12%29.90%26.71%27.00%26.33%
30021.53%24.66%24.35%27.57%29.80%29.69%27.67%28.48%27.33%
40025.50%24.03%29.59%31.18%28.76%29.29%
50023.41%25.60%31.07%31.50%28.90%29.00%
60022.47%21.84%29.37%28.42%28.81%27.57%
70022.88%21.84%30.54%30.12%29.33%28.48%
80024.87%23.72%31.07%31.60%29.52%29.43%
90023.20%24.97%30.33%31.60%29.43%29.86%
100024.66%24.87%30.75%31.71%28.81%29.52%