sft_config.json

{
  "base_model": "models/cheddar-2.0-PRETRAINED",
  "dataset": "data/ultrachat",
  "epochs": 0.25,
  "sequence_length": 2048,
  "micro_batch_size": 1,
  "batch_tokens": 131072,
  "gradient_accumulation_steps": 64,
  "optimizer_steps": 494,
  "adamw_lr": 1.5e-05,
  "muon_lr": 0.0005,
  "weight_decay": 0.01,
  "warmup_fraction": 0.03,
  "decay_fraction": 0.2,
  "min_lr_ratio": 0.05,
  "grad_clip": 1.0,
  "tokens_trained": 64653607
}