rulm

1
{
2
    "trainer": {
3
        "evaluation_strategy": "steps",
4
        "per_device_train_batch_size": 2,
5
        "per_device_eval_batch_size": 2,
6
        "gradient_accumulation_steps": 16,
7
        "eval_steps": 150,
8
        "save_steps": 150,
9
        "logging_steps": 10,
10
        "learning_rate": 0.00005,
11
        "num_train_epochs": 3,
12
        "lr_scheduler_type": "cosine",
13
        "warmup_steps": 100,
14
        "fp16": false,
15
        "bf16": true,
16
        "gradient_checkpointing": false,
17
        "torch_compile": false,
18
        "optim": "adamw_torch",
19
        "half_precision_backend": "auto",
20
        "fp16_opt_level": "O2"
21
    },
22
    "lora": {
23
        "r": 8,
24
        "lora_alpha": 32,
25
        "lora_dropout": 0.1,
26
        "inference_mode": false,
27
        "task_type": "SEQ_2_SEQ_LM"
28
    },
29
    "load_in_8bit": true,
30
    "model_name": "bigscience/mt0-xl",
31
    "model_type": "seq2seq",
32
    "max_source_tokens_count": 256,
33
    "max_target_tokens_count": 256
34
}
35

36