README.md

---
license: mit
---
## Introduction
A paraphrase generation model that *attempts* to give you some control over the output text using natural language. You can do this by prepending the following to the input text:


```
Paraphrase: {distance_keyword} changes, {word_length_keyword} input.
```

### distance_keyword
This tells the model how much to change the input text. There are four options: 
1. small
2. medium
3. large
4. gigantic

### word_length_keyword:
Tells the model how long to make the output text relative to the input. There are three options:
1. reduce
2. match
3. expand

If you only want to paraphrase and don't necessarily care about the specifics of the output, you can also prepend "Paraphrase: " alone or skip the prepending all together and just input the text you wish to paraphrase.

## How to use:

Initializing model using GPU and Bfloat16 precision:

```python
from transformers import pipeline 
from torch import bfloat16

para_gen = pipeline('text2text-generation', model="imjeffhi/paraphrase_generator", tokenizer="imjeffhi/paraphrase_generator", device=0, torch_dtype=bfloat16)
```

Calling model: 

```python
options_phrase = "Paraphrase: large changes, match input."
input_text = "A paraphrase is a restatement of the meaning of a text or passage using other words."
output = para_gen(f"{options_phrase} {input_text}", do_sample=True, top_k=10, num_return_sequences=5)
```
Output: 
```python
[{'generated_text': 'A paraphrase is a modification of the meaning or expression of a text or passage by using other words.'},
 {'generated_text': 'A paraphrase is a continuation of the meaning of a text or a passage using other words.'},
 {'generated_text': 'A paraphrase is the restatement of the meaning of a text or other passage containing other words.'},
 {'generated_text': 'The paraphrase is a repetition of the meanings of a text or passage using other words.'},
 {'generated_text': 'A paraphrase is a continuation of a sentence or passage by using other words.'}]
```