# ATC sweep winner; see doc/results/optimizer-tuning.md.
data:
  revision: 1588ec454efa1a09f29cd18ddd04fe05fc8653a2
  tokenizer_revision: 59f6f375b26bde864a6ca194a9a3044570490064
decentralized:
  num_models: 4
  topology: one_peer_exponential
  scheme: atc
training:
  batch_tokens: 131072
  micro_batch_size: 32
  checkpoint_policy: final
runtime:
  output_dir: runs/packed4-20m-atc
optimizer:
  lr: 0.0056
  beta1: 0.9
  beta2: 0.99
  eps: 1.0e-08
  weight_decay: 0.1
  grad_clip: 1.0
