🦠 Mutate

A library to synthesize text datasets using Large Language Models (LLM). Mutate reads through the examples in the dataset and generates similar examples using auto generated few shot prompts.

1. Installation

pip install mutate-nlp

or

pip install git+https://github.com/infinitylogesh/mutate

2. Usage

2.1 Synthesize text data from local csv files

from mutate import pipeline

pipe = pipeline("text-classification-synthesis",
                model="EleutherAI/gpt-neo-125M",
                device=1)

task_desc = "Each item in the following contains movie reviews and corresponding sentiments. Possible sentimets are neg and pos"


# returns a python generator
text_synth_gen = pipe("csv",
                    data_files=["local/path/sentiment_classfication.csv"],
                    task_desc=task_desc,
                    text_column="text",
                    label_column="label",
                    text_column_alias="Comment",
                    label_column_alias="sentiment",
                    shot_count=5,
                    class_names=["pos","neg"])

#Loop through the generator to synthesize examples by class
for synthesized_examples  in text_synth_gen:
    print(synthesized_examples)

{
    "text": ["The story was very dull and was a waste of my time. This was not a film I would ever watch. The acting was bad. I was bored. There were no surprises. They showed one dinosaur,",
    "I did not like this film. It was a slow and boring film, it didn't seem to have any plot, there was nothing to it. The only good part was the ending, I just felt that the film should have ended more abruptly."]
    "label":["neg","neg"]
}

{
    "text":["The Bell witch is one of the most interesting, yet disturbing films of recent years. It’s an odd and unique look at a very real, but very dark issue. With its mixture of horror, fantasy and fantasy adventure, this film is as much a horror film as a fantasy film. And it‘s worth your time. While the movie has its flaws, it is worth watching and if you are a fan of a good fantasy or horror story, you will not be disappointed."],
    "label":["pos"]
}

# and so on .....

2.2 Synthesize text data from 🤗 datasets

Under the hood Mutate uses the wonderful 🤗 datasets library for dataset processing, So it supports 🤗 datasets out of the box.

from mutate import pipeline

pipe = pipeline("text-classification-synthesis",
                model="EleutherAI/gpt-neo-2.7B",
                device=1)

task_desc = "Each item in the following contains customer service queries expressing the mentioned intent"

synthesizerGen = pipe("banking77",
                    task_desc=task_desc,
                    text_column="text",
                    label_column="label",
                    # if the `text_column` doesn't have a meaningful value
                    text_column_alias="Queries",
                    label_column_alias="Intent", # if the `label_column` doesn't have a meaningful value
                    shot_count=5,
                    dataset_args=["en"])


for exp in synthesizerGen:
    print(exp)

{"text":["How can i know if my account has been activated? (This is the one that I am confused about)",
         "Thanks! My card activated"],
"label":["activate_my_card",
         "activate_my_card"]
}

{
"text": ["How do i activate this new one? Is it possible?",
         "what is the activation process for this card?"],
"label":["activate_my_card",
         "activate_my_card"]
}

# and so on .....

2.3 I am feeling lucky : Infinetly loop through the dataset to generate examples indefinetly

Caution: Infinetly looping through the dataset has a higher chance of duplicate examples to be generated.

from mutate import pipeline

pipe = pipeline("text-classification-synthesis",
                model="EleutherAI/gpt-neo-2.7B",
                device=1)

task_desc = "Each item in the following contains movie reviews and corresponding sentiments. Possible sentimets are neg and pos"


# returns a python generator
text_synth_gen = pipe("csv",
                    data_files=["local/path/sentiment_classfication.csv"],
                    task_desc=task_desc,
                    text_column="text",
                    label_column="label",
                    text_column_alias="Comment",
                    label_column_alias="sentiment",
                    class_names=["pos","neg"],
                    # Flag to generate indefinite examples
                    infinite_loop=True)

#Infinite loop
for exp in synthesizerGen:
    print(exp)

3. Support

3.1 Currently supports

• Text classification dataset synthesis : Few Shot text data synsthesize for text classification datasets using Causal LLMs ( GPT like )

3.2 Roadmap:

• Other types of text Dataset synthesis - NER , sentence pairs etc
• Finetuning support for better quality generation
• Pseudo labelling

4. Credit

• EleutherAI for democratizing Large LMs.
• This library uses 🤗 Datasets and 🤗 Transformers for processing datasets and models.

5. References

The Idea of generating examples from Large Language Model is inspired by the works below,

• A Few More Examples May Be Worth Billions of Parameters by Yuval Kirstain, Patrick Lewis, Sebastian Riedel, Omer Levy
• GPT3Mix: Leveraging Large-scale Language Models for Text Augmentation by Kang Min Yoo, Dongju Park, Jaewook Kang, Sang-Woo Lee, Woomyeong Park
• Data Augmentation using Pre-trained Transformer Models by Varun Kumar, Ashutosh Choudhary, Eunah Cho