{
  "mode": "full_finetune",
  "target_hardware": "8x NVIDIA H200",
  "model_path": "/home/work/.projects/LLM-OS-Models/long-gdn/runs/gdn1_32k_balanced_recovery_1b_bs10_ft/checkpoint-200",
  "tokenizer_path": "/home/work/.projects/LLM-OS-Models/long-gdn/runs/gdn1_32k_balanced_recovery_1b_bs10_ft/final",
  "manifest": "/home/work/.projects/LLM-OS-Models/long-gdn/configs/gdn1_memory_mix_32k_anchor_recovery.json",
  "output_dir": "/home/work/.projects/LLM-OS-Models/long-gdn/runs/gdn1_32k_anchor_from_balanced200_1b_bs10_ft",
  "seq_len": 32768,
  "per_device_batch_size": 10,
  "grad_accum": 1,
  "world_size_assumed": 8,
  "global_tokens_per_step": 2621440,
  "max_steps": 382,
  "target_tokens": 1001390080,
  "learning_rate": 3e-06,
  "warmup_steps": 25,
  "weight_decay": 0.1,
  "optim": "adamw_torch_fused",
  "bf16": true,
  "gradient_checkpointing": true,
  "override_max_position_embeddings": null,
  "notes": [
    "No LoRA/PEFT adapters are used.",
    "All model parameters remain trainable unless the model implementation freezes them.",
    "Dry-run writes this plan without loading model weights."
  ]
}