init
Browse files
README.md
ADDED
@@ -0,0 +1,54 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Cheapity3 π·
|
2 |
+
GPT3-like T5 model trained to generate text in multiple languages.
|
3 |
+
|
4 |
+
## Motivation
|
5 |
+
- GPT models are expensive run.
|
6 |
+
- GPT models are monolingual.
|
7 |
+
|
8 |
+
## Solution
|
9 |
+
- Maybe, Small Models aren't Terrible (*SMarT*)
|
10 |
+
- Plus, they are cheaper to run.
|
11 |
+
|
12 |
+
I fine-tuned T5 on multiple languages (π¬π§ English, π©πͺ German, π«π· French) and multiple academic text snippets from various
|
13 |
+
domains like tech, law, finance and science etc. to generate text, just like GPT models do.
|
14 |
+
|
15 |
+
## Usage
|
16 |
+
- Provide some text e.g `"Italy, officially the Italian Republic is a country consisting of"`
|
17 |
+
- Tell Cheapity3 how many words you want to generate e.g `15` -- π Yes, you can control the length.
|
18 |
+
- Cheapity3 reads your text and generates a continuation containing approximately 15 words.
|
19 |
+
|
20 |
+
```python
|
21 |
+
from transformers import AutoTokenizer, AutoModelWithLMHead
|
22 |
+
|
23 |
+
tokenizer = AutoTokenizer.from_pretrained("flexudy/cheapity3")
|
24 |
+
|
25 |
+
model = AutoModelWithLMHead.from_pretrained("flexudy/cheapity3")
|
26 |
+
|
27 |
+
input_text = "guess: Italy, officially the Italian Republic is a country consisting of { _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ }" # 15 words
|
28 |
+
|
29 |
+
inputs = tokenizer.encode(input_text, return_tensors="pt", truncation=True, max_length=512)
|
30 |
+
|
31 |
+
input_ids = inputs["input_ids"]
|
32 |
+
|
33 |
+
attention_mask = inputs["attention_mask"]
|
34 |
+
|
35 |
+
outputs = model.generate(
|
36 |
+
input_ids=input_ids,
|
37 |
+
attention_mask=attention_mask,
|
38 |
+
max_length=128,
|
39 |
+
do_sample=True,
|
40 |
+
early_stopping=True,
|
41 |
+
num_return_sequences=4,
|
42 |
+
repetition_penalty=2.5
|
43 |
+
)
|
44 |
+
|
45 |
+
for i in range(4):
|
46 |
+
print(tokenizer.decode(outputs[i], skip_special_tokens=True, clean_up_tokenization_spaces=True))
|
47 |
+
|
48 |
+
# >
|
49 |
+
# >
|
50 |
+
# >
|
51 |
+
# >
|
52 |
+
```
|
53 |
+
|
54 |
+
#
|