{
  "model_name_or_path": "meta-llama/Llama-3.1-8B-Instruct",
  "output_dir": "./outputs/WorldDisasterLM-8B",
  "num_train_epochs": 3,
  "per_device_train_batch_size": 4,
  "per_device_eval_batch_size": 4,
  "gradient_accumulation_steps": 4,
  "gradient_checkpointing": true,
  "learning_rate": 0.0002,
  "lr_scheduler_type": "cosine",
  "warmup_ratio": 0.03,
  "weight_decay": 0.001,
  "max_grad_norm": 0.3,
  "optim": "paged_adamw_32bit",
  "fp16": false,
  "bf16": true,
  "max_seq_length": 4096,
  "packing": true,
  "lora_r": 16,
  "lora_alpha": 32,
  "lora_dropout": 0.05,
  "lora_target_modules": [
    "q_proj",
    "k_proj",
    "v_proj",
    "o_proj",
    "gate_proj",
    "up_proj",
    "down_proj"
  ],
  "use_4bit": true,
  "bnb_4bit_quant_type": "nf4",
  "bnb_4bit_compute_dtype": "bfloat16",
  "use_nested_quant": true,
  "save_steps": 100,
  "logging_steps": 25,
  "evaluation_strategy": "steps",
  "eval_steps": 100,
  "save_total_limit": 3,
  "load_best_model_at_end": true,
  "metric_for_best_model": "eval_loss",
  "dataloader_num_workers": 4,
  "seed": 42,
  "report_to": [
    "tensorboard"
  ],
  "dataset_sources": [
    "ReliefWeb",
    "USGS",
    "GDACS",
    "NOAA",
    "OpenFEMA",
    "WHO"
  ],
  "dataset_size": "88+ live records → 711+ instruction samples per run",
  "languages": [
    "en",
    "ne",
    "es",
    "fr",
    "ar",
    "hi",
    "te",
    "zh",
    "ja",
    "ko",
    "pt"
  ],
  "language_names": [
    "English",
    "Nepali (नेपाली)",
    "Spanish",
    "French",
    "Arabic",
    "Hindi",
    "Telugu",
    "Chinese",
    "Japanese",
    "Korean",
    "Portuguese"
  ],
  "training_status": "PENDING — weights not yet generated. Run: python train.py"
}