The evaluated SST checkpoint includes 167.98B source-stage tokens followed by 222.55B continuation tokens, for 390.54B cumulative tokens; continuation jointly trains both towers in BF16 and retains optimizer state for Prefiller/shared parameters while initializing fresh Decoder state.

The evaluated SST checkpoint includes 167.98B source-stage tokens followed by 222.55B continuation tokens, for 390.54B cumulative tokens; continuation jointly trains both towers in BF16 and retains optimizer state for Prefiller/shared parameters while initializing fresh Decoder state. · CiteArk

测量结果

已评估 0/20 项

获得支持 0有争议或冲突 0尚未评估 20

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
model: source stage of SST	167.98 billion tokens	—	尚未评估
model: 67B SST continuation	222.55 billion tokens	—	尚未评估
model: 67B SST	390.54 billion tokens	—	尚未评估
	13,350 updates	—	尚未评估
	17,687 updates	—	尚未评估
	31,037 updates	—	尚未评估
gpu_type: NVIDIA H800	128 GPUs	—	尚未评估
	4,096 tokens	—	尚未评估
	3,072 sequences	—	尚未评估
models: source, 67B SST, and 63B classic	12,582,912 tokens	—	尚未评估
model: 47B classic	14,680,064 tokens	—	尚未评估
	6 steps	—	尚未评估

The evaluated SST checkpoint includes 167.98B source-stage tokens followed by 222.55B continuation tokens, for 390.54B cumulative tokens; continuation jointly trains both towers in BF16 and retains optimizer state for Prefiller/shared parameters while initializing fresh Decoder state.

报告指标与观测值

Assessment（0）

关联运行（0）

测量结果