Training config
Browse files- training_config.json +53 -0
training_config.json
ADDED
|
@@ -0,0 +1,53 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 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": 5,
|
| 10 |
+
"learning_rate": 0.003,
|
| 11 |
+
"num_train_epochs": 5,
|
| 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 |
+
"deepspeed": {
|
| 23 |
+
"bf16": {
|
| 24 |
+
"enabled": true
|
| 25 |
+
},
|
| 26 |
+
"optimizer": {
|
| 27 |
+
"type": "AdamW",
|
| 28 |
+
"params": {
|
| 29 |
+
"lr": "auto",
|
| 30 |
+
"betas": "auto",
|
| 31 |
+
"eps": "auto",
|
| 32 |
+
"weight_decay": "auto"
|
| 33 |
+
}
|
| 34 |
+
},
|
| 35 |
+
"zero_optimization": {
|
| 36 |
+
"stage": 2,
|
| 37 |
+
"offload_optimizer": {
|
| 38 |
+
"device": "cpu",
|
| 39 |
+
"pin_memory": true
|
| 40 |
+
},
|
| 41 |
+
"overlap_comm": true,
|
| 42 |
+
"round_robin_gradients": true
|
| 43 |
+
},
|
| 44 |
+
"train_batch_size": "auto",
|
| 45 |
+
"gradient_accumulation_steps": "auto"
|
| 46 |
+
},
|
| 47 |
+
"model_name": "ai-forever/FRED-T5-1.7B",
|
| 48 |
+
"templates_path": "ru_alpaca_seq2seq_template.json",
|
| 49 |
+
"model_type": "seq2seq",
|
| 50 |
+
"max_source_tokens_count": 512,
|
| 51 |
+
"max_target_tokens_count": 512
|
| 52 |
+
}
|
| 53 |
+
|