{
  "checkpoint": "ckpt_13.pt",
  "model": "GPT2-S",
  "training_step": 10000,
  "instability_type": "Slow divergence",
  "learning_rate": "6e-4",
  "decay": "0.1",
  "warm": "2000",
  "data_type": "FP8_with_FP8_head"
}