Upload finetuning_config.yaml with huggingface_hub
Browse files- finetuning_config.yaml +42 -0
finetuning_config.yaml
ADDED
@@ -0,0 +1,42 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
alpha: 0.25
|
2 |
+
base_model: meta-llama/Llama-3.2-1B-Instruct
|
3 |
+
custom_name: d4-a0.25
|
4 |
+
dtype: bfloat16
|
5 |
+
lambdas:
|
6 |
+
- 1.0
|
7 |
+
- 1.0
|
8 |
+
- 1.0
|
9 |
+
lora_config: null
|
10 |
+
metric: null
|
11 |
+
original_datasets:
|
12 |
+
- !!python/object/apply:finetuning.dataset.DatasetType
|
13 |
+
- AlpacaGPT4
|
14 |
+
- !!python/object/apply:finetuning.dataset.DatasetType
|
15 |
+
- OpenWebText
|
16 |
+
proportions:
|
17 |
+
- 0.1
|
18 |
+
- 0.1
|
19 |
+
training_args:
|
20 |
+
bf16: false
|
21 |
+
do_train: true
|
22 |
+
fp16: false
|
23 |
+
gradient_accumulation_steps: 8
|
24 |
+
gradient_checkpointing: false
|
25 |
+
hub_strategy: all_checkpoints
|
26 |
+
learning_rate: 2.0e-05
|
27 |
+
logging_steps: 10
|
28 |
+
lr_scheduler_type: cosine
|
29 |
+
max_steps: 2500
|
30 |
+
num_train_epochs: 1
|
31 |
+
optim: adafactor
|
32 |
+
output_dir: Grogros/dmWM-llama-3.2-1B-Instruct-HarmData-Al4-OWT-d4-a0.25
|
33 |
+
overwrite_output_dir: true
|
34 |
+
per_device_train_batch_size: 4
|
35 |
+
push_to_hub: true
|
36 |
+
save_steps: 500
|
37 |
+
save_strategy: steps
|
38 |
+
warmup_ratio: 0.1
|
39 |
+
watermark_datasets:
|
40 |
+
- !!python/object/apply:finetuning.dataset.DatasetType
|
41 |
+
- HarmData
|
42 |
+
watermark_eval_config: []
|