sedrickkeh
commited on
Commit
•
e116efc
1
Parent(s):
85e401c
Upload configs.yaml with huggingface_hub
Browse files- configs.yaml +37 -0
configs.yaml
ADDED
@@ -0,0 +1,37 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
bf16: true
|
2 |
+
cutoff_len: 512
|
3 |
+
dataset: llamafactory/alpaca_en
|
4 |
+
dataset_dir: ONLINE
|
5 |
+
ddp_timeout: 180000000
|
6 |
+
deepspeed: dcft/train/zero3.json
|
7 |
+
do_train: true
|
8 |
+
enable_liger_kernel: true
|
9 |
+
eval_strategy: epoch
|
10 |
+
finetuning_type: full
|
11 |
+
formatting: alpaca
|
12 |
+
global_batch_size: 512
|
13 |
+
gradient_accumulation_steps: 8
|
14 |
+
gradient_checkpointing: true
|
15 |
+
hub_model_id: mlfoundations-dev/test_run_mini
|
16 |
+
learning_rate: 2.0e-05
|
17 |
+
logging_steps: 10
|
18 |
+
lr_scheduler_type: cosine
|
19 |
+
max_steps: 3
|
20 |
+
model_name_or_path: meta-llama/Llama-3.2-1B
|
21 |
+
neat_packing: true
|
22 |
+
output_dir: /data4/dcft/experiments/train/checkpoints/test_run_mini
|
23 |
+
overwrite_cache: true
|
24 |
+
overwrite_output_dir: true
|
25 |
+
packing: true
|
26 |
+
per_device_train_batch_size: 8
|
27 |
+
plot_loss: true
|
28 |
+
preprocessing_num_workers: 16
|
29 |
+
push_to_db: false
|
30 |
+
push_to_hub: true
|
31 |
+
report_to: wandb
|
32 |
+
run_name: test_run_mini
|
33 |
+
save_strategy: epoch
|
34 |
+
stage: sft
|
35 |
+
template: alpaca
|
36 |
+
val_size: 0.05
|
37 |
+
warmup_ratio: 0.1
|