Commit
•
054955a
1
Parent(s):
86414a1
Create README.md
Browse files
README.md
ADDED
@@ -0,0 +1,27 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
This model is a finetuned version run by the finetune.py in github repository tolen/alpaca-lora with parameters:
|
2 |
+
|
3 |
+
# model/data params
|
4 |
+
base_model: str = "yahma/llama-7b-hf",
|
5 |
+
data_path: str = "", # dataset see repository https://github.com/ddzipp/AutoAudit_LLM_Dataset/tree/v0.0.1
|
6 |
+
output_dir: str = "./autoaudit_20230703_attempt1",
|
7 |
+
# training hyperparams
|
8 |
+
batch_size: int = 4,
|
9 |
+
micro_batch_size: int = 1,
|
10 |
+
num_epochs: int = 14,
|
11 |
+
learning_rate: float = 3e-4,
|
12 |
+
cutoff_len: int = 512,
|
13 |
+
val_set_size: int = 400,
|
14 |
+
# lora hyperparams
|
15 |
+
lora_r: int = 16,
|
16 |
+
lora_alpha: int = 16,
|
17 |
+
lora_dropout: float = 0.05,
|
18 |
+
lora_target_modules: List[str] = [
|
19 |
+
"q_proj",
|
20 |
+
"k_proj",
|
21 |
+
"v_proj",
|
22 |
+
"o_proj"
|
23 |
+
],
|
24 |
+
# llm hyperparams
|
25 |
+
train_on_inputs: bool = True, # if False, masks out inputs in loss
|
26 |
+
add_eos_token: bool = False,
|
27 |
+
group_by_length: bool = False, # faster, but produces an odd training loss curve
|