The compared models use matched sparse-expert and hybrid-attention patterns but different widths, heads, and body sizes; in SST only the Prefiller contains K/V projections and K normalization, while both towers have independent Transformer-block parameters.

The compared models use matched sparse-expert and hybrid-attention patterns but different widths, heads, and body sizes; in SST only the Prefiller contains K/V projections and K normalization, while both towers have independent Transformer-block parameters. · CiteArk

测量结果

已评估 0/26 项

获得支持 0有争议或冲突 0尚未评估 26

保留同一结论下的完整结果组。已有评估排在前面；单项判断不替代整组结论。展开行查看条件、来源与历次证据。

每行是一项测量，分别显示论文报告、最近可用实测和累计证据状态。
指标与条件	论文报告	最近可用实测	证据判断
models: source and each SST tower	4,608 dimensions	—	尚未评估
model: 47B classic	5,120 dimensions	—	尚未评估
model: 63B classic	5,632 dimensions	—	尚未评估
models: source and each SST tower	576 dimensions	—	尚未评估
models: source and each SST tower	576 dimensions	—	尚未评估
model: 47B classic	640 dimensions	—	尚未评估
model: 47B classic	640 dimensions	—	尚未评估
model: 63B classic	704 dimensions	—	尚未评估
model: 63B classic	704 dimensions	—	尚未评估
models: source, SST, 47B classic, 63B classic	128 dimensions	—	尚未评估
models: source, SST, 47B classic, 63B classic	64 dimensions	—	尚未评估
models: source, SST, 47B classic, 63B classic	128 dimensions	—	尚未评估

The compared models use matched sparse-expert and hybrid-attention patterns but different widths, heads, and body sizes; in SST only the Prefiller contains K/V projections and K normalization, while both towers have independent Transformer-block parameters.

报告指标与观测值

Assessment（0）

关联运行（0）

测量结果