Add: Configuration file for training
Browse files- transformers_config.json +72 -0
transformers_config.json
ADDED
|
@@ -0,0 +1,72 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"model_config": {
|
| 3 |
+
"model_name_or_path": "unsloth/DeepSeek-R1-Distill-Qwen-14B-bnb-4bit",
|
| 4 |
+
"use_cache": false,
|
| 5 |
+
"rope_scaling": {
|
| 6 |
+
"type": "dynamic",
|
| 7 |
+
"factor": 2.0
|
| 8 |
+
}
|
| 9 |
+
},
|
| 10 |
+
"training_config": {
|
| 11 |
+
"num_train_epochs": 3,
|
| 12 |
+
"per_device_train_batch_size": 2,
|
| 13 |
+
"gradient_accumulation_steps": 4,
|
| 14 |
+
"learning_rate": 2e-5,
|
| 15 |
+
"lr_scheduler_type": "cosine",
|
| 16 |
+
"warmup_ratio": 0.03,
|
| 17 |
+
"weight_decay": 0.01,
|
| 18 |
+
"optim": "adamw_torch",
|
| 19 |
+
"max_grad_norm": 0.3,
|
| 20 |
+
"max_seq_length": 2048,
|
| 21 |
+
"logging_steps": 10,
|
| 22 |
+
"save_steps": 200,
|
| 23 |
+
"save_total_limit": 3,
|
| 24 |
+
"evaluation_strategy": "steps",
|
| 25 |
+
"eval_steps": 200,
|
| 26 |
+
"load_best_model_at_end": true,
|
| 27 |
+
"output_dir": "fine_tuned_model",
|
| 28 |
+
"disable_tqdm": false,
|
| 29 |
+
"report_to": ["tensorboard"],
|
| 30 |
+
"logging_first_step": true
|
| 31 |
+
},
|
| 32 |
+
"hardware_config": {
|
| 33 |
+
"fp16": true,
|
| 34 |
+
"bf16": false,
|
| 35 |
+
"gradient_checkpointing": true,
|
| 36 |
+
"device_map": "auto",
|
| 37 |
+
"use_flash_attention": false,
|
| 38 |
+
"attn_implementation": "eager"
|
| 39 |
+
},
|
| 40 |
+
"quantization_config": {
|
| 41 |
+
"load_in_4bit": true,
|
| 42 |
+
"bnb_4bit_compute_dtype": "float16",
|
| 43 |
+
"bnb_4bit_quant_type": "nf4",
|
| 44 |
+
"bnb_4bit_use_double_quant": true
|
| 45 |
+
},
|
| 46 |
+
"lora_config": {
|
| 47 |
+
"r": 16,
|
| 48 |
+
"lora_alpha": 32,
|
| 49 |
+
"lora_dropout": 0.05,
|
| 50 |
+
"bias": "none",
|
| 51 |
+
"target_modules": [
|
| 52 |
+
"q_proj",
|
| 53 |
+
"k_proj",
|
| 54 |
+
"v_proj",
|
| 55 |
+
"o_proj",
|
| 56 |
+
"gate_proj",
|
| 57 |
+
"up_proj",
|
| 58 |
+
"down_proj"
|
| 59 |
+
]
|
| 60 |
+
},
|
| 61 |
+
"dataset_config": {
|
| 62 |
+
"sort_by_field": "prompt_number",
|
| 63 |
+
"sort_direction": "ascending",
|
| 64 |
+
"max_tokens": 2048,
|
| 65 |
+
"text_field": "conversations",
|
| 66 |
+
"shuffle_seed": 42,
|
| 67 |
+
"training_phase_only": true,
|
| 68 |
+
"pre_tokenized": true,
|
| 69 |
+
"input_ids_field": "input_ids",
|
| 70 |
+
"skip_tokenization": true
|
| 71 |
+
}
|
| 72 |
+
}
|