LoGo: Token-Level Dynamic Local-Global Attention
尚未评估计划受阻研究发现claim-scaling
LoGo preserves the reported scaling behavior of full-attention Transformers from 200M through 3.3B parameters, matching or improving the baseline on the listed language-modeling and commonsense metrics.
来源：paper-fixed:Table 1, p. 7
报告指标与观测值
train_loss
m-scaling-33b-logo-train-loss
论文报告 1.996 loss
实际观测 — loss
wikitext_perplexity
m-scaling-33b-logo-wikitext
论文报告 14.785 perplexity
实际观测 — perplexity
commonsense_reasoning_average
m-scaling-33b-logo-csr
论文报告 58.26 percentage_points
实际观测 — percentage_points
Assessment（0）
这条 Claim 暂无已发布的不可变 Assessment。
关联运行（0）
这条 Claim 暂无关联执行。