Training Stability Analysis on Llama-2 1.3B (train)
0Number of SpikesSPAM
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| SPAMTotal Steps=9K2025.02 | 0 | 0 | |
| AdaGCTotal Steps=9K2025.02 | 0 | 0 | |
| WeSaR-GlobalGCTotal Steps=9K2025.02 | 2 | 2.22 | |
| ScaledEmbed-GlobalGCTotal Steps=9K2025.02 | 10 | 11.11 |