metadata/ckpt_139.json

{
  "checkpoint": "ckpt_139.pt",
  "model": "GPT2-L",
  "training_step": 12000,
  "instability_type": "Slow divergence",
  "learning_rate": "1e-3",
  "decay": "0.0",
  "warm": "2000",
  "data_type": "BF16"
}