sapinedamo
commited on
Commit
•
9234eb0
1
Parent(s):
5dbd04e
Create README.md
Browse files
README.md
ADDED
@@ -0,0 +1,65 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
datasets:
|
3 |
+
- bertin-project/alpaca-spanish
|
4 |
+
language:
|
5 |
+
- es
|
6 |
+
license: apache-2.0
|
7 |
+
---
|
8 |
+
|
9 |
+
<div style="text-align:center;width:350px;height:350px;">
|
10 |
+
<img src="https://huggingface.co/hackathon-somos-nlp-2023/salsapaca-native/resolve/main/Alpaca2.png" alt="SAlsapaca logo"">
|
11 |
+
</div>
|
12 |
+
|
13 |
+
|
14 |
+
|
15 |
+
# SAlsapaca: Spanish + Salsa + Alpaca-Native (WIP)
|
16 |
+
|
17 |
+
|
18 |
+
## Adapter Description
|
19 |
+
This adapter was created with the [PEFT](https://github.com/huggingface/peft) library and allowed the base model [chavinlo/alpaca-native](https://huggingface.co/chavinlo/alpaca-native) to be fine-tuned on the [Spanish Alpaca Dataset](https://huggingface.co/datasets/bertin-project/alpaca-spanish) by using the method *LoRA*.
|
20 |
+
|
21 |
+
|
22 |
+
## How to use
|
23 |
+
```py
|
24 |
+
import torch
|
25 |
+
from peft import PeftModel, PeftConfig
|
26 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer
|
27 |
+
|
28 |
+
peft_model_id = "hackathon-somos-nlp-2023/salsapaca-native"
|
29 |
+
config = PeftConfig.from_pretrained(peft_model_id)
|
30 |
+
model = AutoModelForCausalLM.from_pretrained(config.base_model_name_or_path, return_dict=True, load_in_8bit=True, device_map='auto')
|
31 |
+
tokenizer = AutoTokenizer.from_pretrained(peft_model_id)
|
32 |
+
|
33 |
+
# Load the Lora model
|
34 |
+
model = PeftModel.from_pretrained(model, peft_model_id)
|
35 |
+
|
36 |
+
def gen_conversation(text):
|
37 |
+
text = "<SC>instruction: " + text + "\n "
|
38 |
+
batch = tokenizer(text, return_tensors='pt')
|
39 |
+
with torch.cuda.amp.autocast():
|
40 |
+
output_tokens = model.generate(**batch, max_new_tokens=256, eos_token_id=50258, early_stopping = True, temperature=.9)
|
41 |
+
|
42 |
+
print('\n\n', tokenizer.decode(output_tokens[0], skip_special_tokens=False))
|
43 |
+
|
44 |
+
text = "hola"
|
45 |
+
|
46 |
+
gen_conversation(text)
|
47 |
+
```
|
48 |
+
|
49 |
+
|
50 |
+
## Resources used
|
51 |
+
Google Colab machine with the following specifications
|
52 |
+
<div style="text-align:center;width:550px;height:550px;">
|
53 |
+
<img src="https://huggingface.co/hackathon-somos-nlp-2023/bertin-gpt-j-6B-es-finetuned-salpaca/resolve/main/resource.jpeg" alt="Resource logo">
|
54 |
+
</div>
|
55 |
+
|
56 |
+
## Citation
|
57 |
+
```
|
58 |
+
@misc {hackathon-somos-nlp-2023,
|
59 |
+
author = { {Edison Bejarano, Leonardo Bolaños, Alberto Ceballos, Santiago Pineda, Nicolay Potes} },
|
60 |
+
title = { SAlsapaca },
|
61 |
+
year = 2023,
|
62 |
+
url = { https://huggingface.co/hackathon-somos-nlp-2023/salsapaca-native }
|
63 |
+
publisher = { Hugging Face }
|
64 |
+
}
|
65 |
+
```
|