rulm

1
{
2
    "trainer": {
3
        "evaluation_strategy": "steps",
4
        "per_device_train_batch_size": 1,
5
        "per_device_eval_batch_size": 1,
6
        "gradient_accumulation_steps": 32,
7
        "eval_steps": 80,
8
        "save_steps": 80,
9
        "logging_steps": 10,
10
        "learning_rate": 0.00003,
11
        "num_train_epochs": 3,
12
        "lr_scheduler_type": "cosine",
13
        "warmup_steps": 80,
14
        "fp16": true,
15
        "bf16": false,
16
        "gradient_checkpointing": false,
17
        "torch_compile": false,
18
        "optim": "adamw_torch",
19
        "half_precision_backend": "auto",
20
        "fp16_opt_level": "O2"
21
    },
22
    "model_name": "sberbank-ai/rugpt3large_based_on_gpt2",
23
    "max_source_tokens_count": 256,
24
    "max_target_tokens_count": 256
25
}
26

27