Commit
•
6eae9eb
1
Parent(s):
371a6e8
Upload trainer_config.yaml
Browse files- trainer_config.yaml +28 -0
trainer_config.yaml
ADDED
|
@@ -0,0 +1,28 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
cutoff_len: 4096
|
| 2 |
+
dataset: LangGPT_community,LangGPT_alpaca
|
| 3 |
+
dataset_dir: /datas/wangm/LLM4LangGPT
|
| 4 |
+
do_train: true
|
| 5 |
+
finetuning_type: lora
|
| 6 |
+
flash_attn: auto
|
| 7 |
+
fp16: true
|
| 8 |
+
gradient_accumulation_steps: 8
|
| 9 |
+
learning_rate: 5.0e-05
|
| 10 |
+
logging_steps: 5
|
| 11 |
+
lora_alpha: 16
|
| 12 |
+
lora_dropout: 0
|
| 13 |
+
lora_rank: 8
|
| 14 |
+
lora_target: q_proj,v_proj
|
| 15 |
+
lr_scheduler_type: cosine
|
| 16 |
+
max_grad_norm: 1.0
|
| 17 |
+
max_samples: 100000
|
| 18 |
+
model_name_or_path: Qwen/Qwen1.5-4B-Chat
|
| 19 |
+
num_train_epochs: 10.0
|
| 20 |
+
optim: adamw_torch
|
| 21 |
+
output_dir: /datas/wangm/LLM4LangGPT/output/Qwen1.5-4B-Chat/full_alpaca
|
| 22 |
+
packing: false
|
| 23 |
+
per_device_train_batch_size: 2
|
| 24 |
+
report_to: none
|
| 25 |
+
save_steps: 100
|
| 26 |
+
stage: sft
|
| 27 |
+
template: default
|
| 28 |
+
warmup_steps: 0
|