We encountered two major loss spikes while training K2.
The first loss spike occured after 160 checkpoints and lasted over ~34 checkpoints. We restarted training at checkpoint 160 and training returned to normal.
The second loss spike occured after restarting training to fix the first loss spike at checkpoint 186 and lasted from ~8 checkpoints.
For every spike checkpoint, we also uploaded the corresponding normal checkpoint for easy comparison. You could find different checkpoints in different branches.
We are releasing these checkpoints so others can study this interesting phenomena in large model training.
k2 loss spikes
Purpose
Loss spikes are still a relatively unknown phenomena. By making these spikes and associated training details available, we hope others use these artifacts to further the worlds knowledge on this topic.
The LLM360 Research Suite is a comprehensive set of large language model (LLM) artifacts from Amber, CrystalCoder, and K2 for academic and industry researchers to explore LLM training dynamics. Additional resources can be found at llm360.ai.
Citation
BibTeX:
bibtex
1@misc{
2 title={LLM360-K2-65B: Scaling Up Open and Transparent Language Models},
3 author={The LLM360 Team},
4 year={2024},
5}
