Upload training_config.json with huggingface_hub
Browse files- training_config.json +56 -0
training_config.json
ADDED
|
@@ -0,0 +1,56 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"stage": "sft",
|
| 3 |
+
"do_train": true,
|
| 4 |
+
"finetuning_type": "lora",
|
| 5 |
+
"model": "Qwen/Qwen3-4B-Instruct-2507",
|
| 6 |
+
"dataset": "llamafactory/fiqa",
|
| 7 |
+
"max_samples": null,
|
| 8 |
+
"eval_dataset": null,
|
| 9 |
+
"train_dataset_num_samples": 5500,
|
| 10 |
+
"eval_dataset_num_samples": null,
|
| 11 |
+
"quantization_bit": 4,
|
| 12 |
+
"load_in_4bit": true,
|
| 13 |
+
"load_in_16bit": false,
|
| 14 |
+
"lora_rank": 64,
|
| 15 |
+
"lora_r": 64,
|
| 16 |
+
"lora_alpha": 128,
|
| 17 |
+
"lora_dropout": 0.05,
|
| 18 |
+
"target_modules": [
|
| 19 |
+
"q_proj",
|
| 20 |
+
"k_proj",
|
| 21 |
+
"v_proj",
|
| 22 |
+
"o_proj",
|
| 23 |
+
"gate_proj",
|
| 24 |
+
"up_proj",
|
| 25 |
+
"down_proj"
|
| 26 |
+
],
|
| 27 |
+
"bias": "none",
|
| 28 |
+
"use_gradient_checkpointing": "unsloth",
|
| 29 |
+
"random_state": 3407,
|
| 30 |
+
"learning_rate": 0.00015,
|
| 31 |
+
"lr_scheduler_type": "cosine",
|
| 32 |
+
"warmup_ratio": 0.05,
|
| 33 |
+
"weight_decay": 0.0,
|
| 34 |
+
"adam_beta2": 0.999,
|
| 35 |
+
"num_train_epochs": 4.0,
|
| 36 |
+
"max_steps": -1,
|
| 37 |
+
"per_device_train_batch_size": 4,
|
| 38 |
+
"gradient_accumulation_steps": 4,
|
| 39 |
+
"seq_len": 2048,
|
| 40 |
+
"max_seq_length": 2048,
|
| 41 |
+
"logging_steps": 20,
|
| 42 |
+
"save_strategy": "steps",
|
| 43 |
+
"save_steps": 500,
|
| 44 |
+
"save_total_limit": 2,
|
| 45 |
+
"evaluation_strategy": "no",
|
| 46 |
+
"eval_steps": null,
|
| 47 |
+
"load_best_model_at_end": false,
|
| 48 |
+
"bf16": true,
|
| 49 |
+
"report_to": "none",
|
| 50 |
+
"dataset_num_proc": 4,
|
| 51 |
+
"seed": 3407,
|
| 52 |
+
"output_dir": "/root/jb/personas/finance/fiqa/finance_lora_unsloth_output/20260305_020931",
|
| 53 |
+
"adam_beta1": 0.9,
|
| 54 |
+
"packing": false,
|
| 55 |
+
"dataset_text_field": "text"
|
| 56 |
+
}
|