Create README.md
Browse files
README.md
ADDED
@@ -0,0 +1,59 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
license: apache-2.0
|
3 |
+
---
|
4 |
+
|
5 |
+
this is a demo how fine tune phi-2 model.
|
6 |
+
|
7 |
+
```
|
8 |
+
import torch
|
9 |
+
import datasets
|
10 |
+
from transformers import TrainingArguments, AutoConfig, AutoTokenizer, AutoModelForCausalLM
|
11 |
+
import trl
|
12 |
+
from transformers import BitsAndBytesConfig
|
13 |
+
|
14 |
+
train_dataset = datasets.load_dataset('HuggingFaceTB/cosmopedia-20k', split='train')
|
15 |
+
|
16 |
+
args = TrainingArguments(
|
17 |
+
output_dir="./test-sft",
|
18 |
+
max_steps=20000,
|
19 |
+
per_device_train_batch_size=1,
|
20 |
+
optim="adafactor", report_to="none",
|
21 |
+
)
|
22 |
+
|
23 |
+
model_id = "microsoft/phi-2"
|
24 |
+
tokenizer = AutoTokenizer.from_pretrained(model_id)
|
25 |
+
nf4_config = BitsAndBytesConfig(
|
26 |
+
load_in_4bit=True,
|
27 |
+
bnb_4bit_quant_type="nf4",
|
28 |
+
bnb_4bit_use_double_quant=True,
|
29 |
+
bnb_4bit_compute_dtype=torch.bfloat16
|
30 |
+
)
|
31 |
+
|
32 |
+
model = AutoModelForCausalLM.from_pretrained(model_id, quantization_config=nf4_config,device_map="auto")
|
33 |
+
print(model)
|
34 |
+
|
35 |
+
from peft import LoraConfig
|
36 |
+
|
37 |
+
peft_config = LoraConfig(
|
38 |
+
lora_alpha=16,
|
39 |
+
lora_dropout=0.1,
|
40 |
+
r=64, target_modules=["q_proj", "v_proj", "k_proj", "dense", "lm_head", "fc1", "fc2"],
|
41 |
+
bias="none",
|
42 |
+
task_type="CAUSAL_LM",
|
43 |
+
)
|
44 |
+
model.add_adapter(peft_config)
|
45 |
+
|
46 |
+
trainer = trl.SFTTrainer(
|
47 |
+
model=model,
|
48 |
+
args=args,
|
49 |
+
train_dataset=train_dataset,
|
50 |
+
dataset_text_field='text',
|
51 |
+
max_seq_length=1024
|
52 |
+
)
|
53 |
+
|
54 |
+
trainer.train()
|
55 |
+
|
56 |
+
trainer.model.save_pretrained("sft", dtype=torch.bfloat16)
|
57 |
+
trainer.tokenizer.save_pretrained("sft")
|
58 |
+
|
59 |
+
```
|