sft_argument.json

{
    "model_name_or_path": "bigscience/bloomz-7b1-mt",
    "dataset_name_or_path": "./data",
    "output_dir": "./checkpoints/bloom_sft_ckpts",
    "per_device_train_batch_size": 4,
    "gradient_accumulation_steps": 4,
    "per_device_eval_batch_size": 8,
    "eval_accumulation_steps":16,
    "num_train_epochs": 3,
    "learning_rate": 3e-05,
    "warmup_steps": 30,
    "logging_steps": 1,
    "evaluation_strategy": "epoch",
    "save_strategy": "epoch",
    "src_length": 1024,
    "max_length": 2048,
    "fp16": true,
    "fp16_opt_level": "O2",
    "do_train": true,
    "do_eval": true,
    "disable_tqdm": true,
    "load_best_model_at_end": true,
    "eval_with_do_generation": false,
    "metric_for_best_model": "accuracy",
    "recompute": true,
    "save_total_limit": 1,
    "tensor_parallel_degree": 4,
    "pipeline_parallel_degree": 1,
    "zero_padding": false,
    "use_flash_attention": false
  }