training_metadata.json

{
  "model_name": "microsoft/Phi-3.5-mini-instruct",
  "display_name": "Phi-3.5 (BF16, Liger, Batch96, CkptON)",
  "timestamp": "2026-02-03T19:35:19.489805",
  "training_config": {
    "num_train_epochs": 1,
    "per_device_train_batch_size": 96,
    "gradient_accumulation_steps": 1,
    "learning_rate": 0.0002,
    "warmup_ratio": 0.03,
    "lr_scheduler_type": "cosine",
    "weight_decay": 0.01,
    "max_seq_length": 4096,
    "logging_steps": 1,
    "eval_steps": 50,
    "save_steps": 200,
    "seed": 42,
    "bf16": true,
    "optim": "adamw_torch_fused",
    "dataloader_num_workers": 8,
    "torch_compile": true
  },
  "lora_config": {
    "r": 16,
    "lora_alpha": 32,
    "lora_dropout": 0.05,
    "target_modules": [
      "q_proj",
      "k_proj",
      "v_proj",
      "o_proj",
      "gate_proj",
      "up_proj",
      "down_proj"
    ],
    "bias": "none",
    "task_type": "CAUSAL_LM"
  },
  "train_loss": 0.5981301681586066,
  "train_samples": 172145,
  "val_samples": 9066,
  "train_time_minutes": 147.3416652202606,
  "max_memory_gb": 54.91591787338257,
  "fix_applied": "YAML normalization via PyYAML (2 spaces), packing=False, Native BF16 Training, Batch 96"
}