{
  "optimizer_type": "adamw",
  "muon_lr": 0.0001,
  "adam_lr": 0.0001,
  "weight_decay": 0.01,
  "muon_momentum": 0.95,
  "ns_steps": 5,
  "batch_size": 16,
  "gradient_accumulation_steps": 1,
  "num_epochs": 2,
  "max_grad_norm": 0.5,
  "use_mixed_precision": true,
  "amp_dtype": "bfloat16",
  "max_train_steps": null,
  "dataloader_num_workers": 8,
  "pin_memory": true,
  "dataloader_prefetch_factor": 4,
  "warmup_ratio": 0.1,
  "logging_steps": 25,
  "eval_steps": 100,
  "save_steps": 200,
  "early_stopping_patience": 3,
  "early_stopping_min_delta": 0.0001,
  "early_stopping_monitor": "val_loss",
  "early_stopping_enabled": false,
  "length_grouped_batches": true,
  "max_val_batches": 50,
  "lr_scheduler": "cosine_with_warmup",
  "lr_scheduler_patience": 3,
  "lr_scheduler_factor": 0.5,
  "lr_scheduler_min_lr": 1e-06,
  "output_dir": "/outputs/ablation_budget/csa_mhc_adamw/baseline",
  "experiment_name": "evlms_v2_baseline",
  "skip_nonfinite_batches": true,
  "max_nonfinite_batches": 20,
  "save_epoch_artifacts": true,
  "epoch_artifacts_interval": 1,
  "epoch_gen_samples": 8,
  "gen_max_new_tokens": 320,
  "hf_push_async": true,
  "use_pos_weight": true,
  "pos_weight_max_samples": 512,
  "pos_weight_max": 20.0
}