Create README.md
Browse files
README.md
ADDED
@@ -0,0 +1,37 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
pipeline_tag: multilabel-classification
|
3 |
+
---
|
4 |
+
Multi Label Classification
|
5 |
+
|
6 |
+
Description: Labeling the product comments of the e-commerce data we have created according to the labels we have specified.
|
7 |
+
It returns us a positive or negative value based on the comments made on the product or shipping.
|
8 |
+
If no comment is made, None data regarding the product or shipping will be returned. In this project I used the model "meta-llama/Meta-Llama-3-8B-Instruct"
|
9 |
+
|
10 |
+
Used to quantization method like this ;
|
11 |
+
|
12 |
+
>>
|
13 |
+
import torch
|
14 |
+
#quantizasyon yöntemi 4 bit hassasiyetiyle yapacak.
|
15 |
+
quant_config = BitsAndBytesConfig(
|
16 |
+
load_in_4bit=True, #olabildiğince küçültmeye çalışıyoruz maliyet açısından, ama doğruluğu da düşüyor
|
17 |
+
bnb_4bit_quant_type="nf4", #"fp4" , nf4 daha az yer kaplar daha hızlıdır ama doğruluğu kıyasla daha az
|
18 |
+
bnb_4bit_compute_dtype=torch.float16, #maliyeti de düşürmüş olduk, hızı da artırdık
|
19 |
+
bnb_4bit_use_double_quant=True, # True model increases accuracy but causes it to work more costly.
|
20 |
+
) #np4 saves more memory
|
21 |
+
>>
|
22 |
+
|
23 |
+
|
24 |
+
Here.. This code contains important steps to optimize LLM's memory usage and processing time.
|
25 |
+
It is important to improve the performance and resource utilization of the model.
|
26 |
+
>>
|
27 |
+
#20GB model
|
28 |
+
model = AutoModelForCausalLM.from_pretrained(
|
29 |
+
pretrained_model_name_or_path=base_model, #used "meta-llama/Meta-Llama-3-8B-Instruct"
|
30 |
+
quantization_config=quant_config,
|
31 |
+
device_map={"": 0},
|
32 |
+
#use_auth_token=True
|
33 |
+
|
34 |
+
)
|
35 |
+
model.config.use_cache = False
|
36 |
+
model.config.pretraining_tp = 1
|
37 |
+
>>
|