sapinedamo
commited on
Commit
•
c0701e7
1
Parent(s):
a9b3346
Update README.md
Browse files
README.md
CHANGED
@@ -10,4 +10,33 @@ license: apache-2.0
|
|
10 |
|
11 |
|
12 |
## Adapter Description
|
13 |
-
This adapter was created with the [PEFT](https://github.com/huggingface/peft) library and allowed the base model *bertin-project/bertin-gpt-j-6B* to be fine-tuned on the *Spanish Alpaca Dataset* by using the method *LoRA*.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
10 |
|
11 |
|
12 |
## Adapter Description
|
13 |
+
This adapter was created with the [PEFT](https://github.com/huggingface/peft) library and allowed the base model *bertin-project/bertin-gpt-j-6B* to be fine-tuned on the *Spanish Alpaca Dataset* by using the method *LoRA*.
|
14 |
+
|
15 |
+
|
16 |
+
## How to use
|
17 |
+
```py
|
18 |
+
import torch
|
19 |
+
from peft import PeftModel, PeftConfig
|
20 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer
|
21 |
+
|
22 |
+
peft_model_id = "hackathon-somos-nlp-2023/bertin-gpt-j-6B-es-finetuned-salpaca"
|
23 |
+
config = PeftConfig.from_pretrained(peft_model_id)
|
24 |
+
model = AutoModelForCausalLM.from_pretrained(config.base_model_name_or_path, return_dict=True, load_in_8bit=True, device_map='auto')
|
25 |
+
# tokenizer = AutoTokenizer.from_pretrained(config.base_model_name_or_path)
|
26 |
+
tokenizer = AutoTokenizer.from_pretrained(peft_model_id)
|
27 |
+
|
28 |
+
# Load the Lora model
|
29 |
+
model = PeftModel.from_pretrained(model, peft_model_id)
|
30 |
+
|
31 |
+
def gen_conversation(text):
|
32 |
+
text = "<SC>instruction: " + text + "\n "
|
33 |
+
batch = tokenizer(text, return_tensors='pt')
|
34 |
+
with torch.cuda.amp.autocast():
|
35 |
+
output_tokens = model.generate(**batch, max_new_tokens=256, eos_token_id=50258, early_stopping = True, temperature=.9)
|
36 |
+
|
37 |
+
print('\n\n', tokenizer.decode(output_tokens[0], skip_special_tokens=False))
|
38 |
+
|
39 |
+
text = "hola"
|
40 |
+
|
41 |
+
gen_conversation(text)
|
42 |
+
```
|