{
  "adam_beta1": 0.9,
  "adam_beta2": 0.95,
  "adam_epsilon": 1e-08,
  "cpu_prefetch_factor": 2,
  "cpu_workers": 2,
  "dataset_manifest": "/home/nvidia/llm-data/haiku/tokenized/sft-all-v3/manifest.json",
  "ema_inv_gamma": 1.0,
  "ema_max_decay": 0.9999,
  "ema_power": 0.75,
  "enable_gradient_checkpointing": true,
  "exclude_components": [
    "nemotron-cascade-2"
  ],
  "fp32_master_weights": true,
  "full_shuffle": true,
  "gradient_accumulation_steps": 16,
  "gradient_checkpointing_layers": 22,
  "gradient_clip_norm": 1.0,
  "initial_run_dir": "/media/nvidia/4TB2/llm-runs/senryu/pretrain-v1",
  "initial_weights": "ema",
  "learning_rate": 0.0004,
  "log_every_optimizer_steps": 1,
  "logit_z_loss_coefficient": 0.0001,
  "maximum_checkpoints": 5,
  "maximum_optimizer_steps": null,
  "micro_batch_size": 1,
  "minimum_learning_rate_ratio": 0.1,
  "muon_learning_rate": null,
  "muon_lr_adjustment": "match_rms_adamw",
  "muon_momentum": 0.95,
  "optimizer": "hybrid_muon",
  "output_dir": "/media/nvidia/4TB2/llm-runs/senryu/sft-v1",
  "require_complete_pretraining": false,
  "require_complete_sft_source": false,
  "rewrite_empty_think": true,
  "save_every_optimizer_steps": 100,
  "seed": 20261008,
  "sequence_length": 32768,
  "shuffle_block_sequences": 2048,
  "stream_download_threads": 8,
  "stream_english_share": null,
  "stream_open_languages": null,
  "stream_pairs_per_row": 1,
  "tokenizer": null,
  "unclosed_think": "mask_end",
  "validation_batches": 16,
  "warmup_ratio": 0.0,
  "warmup_steps": 200,
  "weight_decay": 0.1
}