paddlenlp

1
{
2
    "model_name_or_path": "facebook/opt-125m",
3
    "dataset_name_or_path": "./data",
4
    "output_dir": "./checkpoints/opt_lora_ckpts",
5
    "per_device_train_batch_size": 4,
6
    "gradient_accumulation_steps": 4,
7
    "per_device_eval_batch_size": 8,
8
    "eval_accumulation_steps":16,
9
    "num_train_epochs": 3,
10
    "learning_rate": 3e-04,
11
    "warmup_steps": 30,
12
    "logging_steps": 1,
13
    "evaluation_strategy": "epoch",
14
    "save_strategy": "epoch",
15
    "src_length": 1024,
16
    "max_length": 2048,
17
    "fp16": true,
18
    "fp16_opt_level": "O2",
19
    "do_train": true,
20
    "do_eval": true,
21
    "disable_tqdm": true,
22
    "load_best_model_at_end": true,
23
    "eval_with_do_generation": false,
24
    "metric_for_best_model": "accuracy",
25
    "recompute": true,
26
    "save_total_limit": 1,
27
    "tensor_parallel_degree": 1,
28
    "pipeline_parallel_degree": 1,
29
    "lora": true,
30
    "zero_padding": false,
31
    "use_flash_attention": false
32
  }