Dataset Viber is your chill repo for data collection, annotation and vibe checks.

Project description

Dataset Viber

Avoid the hype, check the vibe!

I've cooked up Dataset Viber, a cool set of tools to make your life easier when dealing with data for AI models. Dataset Viber is all about making your data prep journey smooth and fun. It's not for team collaboration or production, nor trying to be all fancy and formal - just a bunch of cool tools to help you collect feedback and do vibe-checks as an AI engineer or lover. Want to see it in action? Just plug it in and start vibing with your data. It's that easy!

CollectorInterface: Lazily collect data of model interactions without human annotation.
AnnotatorInterface: Walk through your data and annotate it with models in the loop.
BulkInterface: Explore your data distribution and annotate in bulk.
Embdedder: Efficiently embed data with ONNX-optimized speeds.

Need any tweaks or want to hear more about a specific tool? Just open an issue or give me a shout!

[!NOTE]

Data is logged to a local CSV or directly to the Hugging Face Hub.

All tools also run in .ipynb notebooks.

Models in the loop through fn_model.

Input data streamers through fn_next_input.

It supports various tasks for text, chat and image modalities.

Import and export from the Hugging Face Hub or CSV files.

[!TIP] Examples can be found in src/dataset_viber/examples.

Installation

You can install the package via pip:

pip install dataset-viber

Or install BulkInterface dependencies:

pip install dataset-viber[bulk]

How are we vibing?

CollectorInterface

Built on top of the gr.Interface and gr.ChatInterface to lazily collect data for interactions automatically.

https://github.com/user-attachments/assets/4ddac8a1-62ab-4b3b-9254-f924f5898075

Hub dataset

CollectorInterface

import gradio as gr
from dataset_viber import CollectorInterface

def calculator(num1, operation, num2):
    if operation == "add":
        return num1 + num2
    elif operation == "subtract":
        return num1 - num2
    elif operation == "multiply":
        return num1 * num2
    elif operation == "divide":
        return num1 / num2

inputs = ["number", gr.Radio(["add", "subtract", "multiply", "divide"]), "number"]
outputs = "number"

interface = CollectorInterface(
    fn=calculator,
    inputs=inputs,
    outputs=outputs,
    csv_logger=False, # True if you want to log to a CSV
    dataset_name="<my_hf_org>/<my_dataset>"
)
interface.launch()

CollectorInterface.from_interface

interface = gr.Interface(
    fn=calculator,
    inputs=inputs,
    outputs=outputs
)
interface = CollectorInterface.from_interface(
   interface=interface,
   csv_logger=False, # True if you want to log to a CSV
   dataset_name="<my_hf_org>/<my_dataset>"
)
interface.launch()

CollectorInterface.from_pipeline

from transformers import pipeline
from dataset_viber import CollectorInterface

pipeline = pipeline("text-classification", model="mrm8488/bert-tiny-finetuned-sms-spam-detection")
interface = CollectorInterface.from_pipeline(
    pipeline=pipeline,
    csv_logger=False, # True if you want to log to a CSV
    dataset_name="<my_hf_org>/<my_dataset>"
)
interface.launch()

AnnotatorInterface

Built on top of the CollectorInterface to collect and annotate data and log it to the Hub.

Text

https://github.com/user-attachments/assets/d1abda66-9972-4c60-89d2-7626f5654f15

Hub dataset

text-classification/multi-label-text-classification

from dataset_viber import AnnotatorInterFace

texts = [
    "Anthony Bourdain was an amazing chef!",
    "Anthony Bourdain was a terrible tv persona!"
]
labels = ["positive", "negative"]

interface = AnnotatorInterFace.for_text_classification(
    texts=texts,
    labels=labels,
    multi_label=False, # True if you have multi-label data
    fn_model=None, # a callable e.g. (function or transformers pipelines) that returns `str`
    fn_next_input=None, # a function that feeds gradio components actively with the next input
    csv_logger=False, # True if you want to log to a CSV
    dataset_name=None # "<my_hf_org>/<my_dataset>" if you want to log to the hub
)
interface.launch()

token-classification

from dataset_viber import AnnotatorInterFace

texts = ["Anthony Bourdain was an amazing chef in New York."]
labels = ["NAME", "LOC"]

interface = AnnotatorInterFace.for_token_classification(
    texts=texts,
    labels=labels,
    fn_model=None, # a callable e.g. (function or transformers pipelines) that returns `str`
    fn_next_input=None, # a function that feeds gradio components actively with the next input
    csv_logger=False, # True if you want to log to a CSV
    dataset_name=None # "<my_hf_org>/<my_dataset>" if you want to log to the hub
)
interface.launch()

extractive-question-answering

from dataset_viber import AnnotatorInterFace

questions = ["Where was Anthony Bourdain located?"]
contexts = ["Anthony Bourdain was an amazing chef in New York."]

interface = AnnotatorInterFace.for_question_answering(
    questions=questions,
    contexts=contexts,
    fn_model=None, # a callable e.g. (function or transformers pipelines) that returns `str`
    fn_next_input=None, # a function that feeds gradio components actively with the next input
    csv_logger=False, # True if you want to log to a CSV
    dataset_name=None # "<my_hf_org>/<my_dataset>" if you want to log to the hub
)
interface.launch()

text-generation/translation/completion

from dataset_viber import AnnotatorInterFace

prompts = ["Tell me something about Anthony Bourdain."]
completions = ["Anthony Michael Bourdain was an American celebrity chef, author, and travel documentarian."]

interface = AnnotatorInterFace.for_text_generation(
    prompts=prompts, # source
    completions=completions, # optional to show initial completion / target
    fn_model=None, # a callable e.g. (function or transformers pipelines) that returns `str`
    fn_next_input=None, # a function that feeds gradio components actively with the next input
    csv_logger=False, # True if you want to log to a CSV
    dataset_name=None # "<my_hf_org>/<my_dataset>" if you want to log to the hub
)
interface.launch()

text-generation-preference

from dataset_viber import AnnotatorInterFace

prompts = ["Tell me something about Anthony Bourdain."]
completions_a = ["Anthony Michael Bourdain was an American celebrity chef, author, and travel documentarian."]
completions_b = ["Anthony Michael Bourdain was an cool guy that knew how to cook."]

interface = AnnotatorInterFace.for_text_generation_preference(
    prompts=prompts,
    completions_a=completions_a,
    completions_b=completions_b,
    fn_model=None, # a callable e.g. (function or transformers pipelines) that returns `str`
    fn_next_input=None, # a function that feeds gradio components actively with the next input
    csv_logger=False, # True if you want to log to a CSV
    dataset_name=None # "<my_hf_org>/<my_dataset>" if you want to log to the hub
)
interface.launch()

Chat and multi-modal chat

https://github.com/user-attachments/assets/fe7f0139-95a3-40e8-bc03-e37667d4f7a9

Hub dataset

[!TIP] I recommend uploading the files files to a cloud storage and using the remote URL to avoid any issues. This can be done using Hugging Face Datasets. As shown in utils. Additionally GradioChatbot shows how to use the chatbot interface for multi-modal.

chat-classification

from dataset_viber import AnnotatorInterFace

prompts = [
    [
        {
            "role": "user",
            "content": "Tell me something about Anthony Bourdain."
        },
        {
            "role": "assistant",
            "content": "Anthony Michael Bourdain was an American celebrity chef, author, and travel documentarian."
        }
    ]
]

interface = AnnotatorInterFace.for_chat_classification(
    prompts=prompts,
    labels=["toxic", "non-toxic"],
    multi_label=False, # True if you have multi-label data
    fn_model=None, # a callable e.g. (function or transformers pipelines) that returns `str`
    fn_next_input=None, # a function that feeds gradio components actively with the next input
    csv_logger=False, # True if you want to log to a CSV
    dataset_name=None # "<my_hf_org>/<my_dataset>" if you want to log to the hub
)
interface.launch()

chat-generation

from dataset_viber import AnnotatorInterFace

prompts = [
    [
        {
            "role": "user",
            "content": "Tell me something about Anthony Bourdain."
        }
    ]
]

completions = [
    "Anthony Michael Bourdain was an American celebrity chef, author, and travel documentarian.",
]

interface = AnnotatorInterFace.for_chat_generation(
    prompts=prompts,
    completions=completions,
    fn_model=None, # a callable e.g. (function or transformers pipelines) that returns `str`
    fn_next_input=None, # a function that feeds gradio components actively with the next input
    csv_logger=False, # True if you want to log to a CSV
    dataset_name=None # "<my_hf_org>/<my_dataset>" if you want to log to the hub
)
interface.launch()

chat-generation-preference

from dataset_viber import AnnotatorInterFace

prompts = [
    [
        {
            "role": "user",
            "content": "Tell me something about Anthony Bourdain."
        }
    ]
]
completions_a = [
    "Anthony Michael Bourdain was an American celebrity chef, author, and travel documentarian.",
]
completions_b = [
    "Anthony Michael Bourdain was an cool guy that knew how to cook."
]

interface = AnnotatorInterFace.for_chat_generation_preference(
    prompts=prompts,
    completions_a=completions_a,
    completions_b=completions_b,
    fn_model=None, # a callable e.g. (function or transformers pipelines) that returns `str`
    fn_next_input=None, # a function that feeds gradio components actively with the next input
    csv_logger=False, # True if you want to log to a CSV
    dataset_name=None # "<my_hf_org>/<my_dataset>" if you want to log to the hub
)
interface.launch()

Image and multi-modal

https://github.com/user-attachments/assets/57d89edf-ae40-4942-a20a-bf8443100b66

Hub dataset

[!TIP] I recommend uploading the files files to a cloud storage and using the remote URL to avoid any issues. This can be done using Hugging Face Datasets. As shown in utils.

image-classification/multi-label-image-classification

from dataset_viber import AnnotatorInterFace

images = [
    "https://upload.wikimedia.org/wikipedia/commons/thumb/a/a5/Anthony_Bourdain_Peabody_2014b.jpg/440px-Anthony_Bourdain_Peabody_2014b.jpg",
    "https://upload.wikimedia.org/wikipedia/commons/8/85/David_Chang_David_Shankbone_2010.jpg"
]
labels = ["anthony-bourdain", "not-anthony-bourdain"]

interface = AnnotatorInterFace.for_image_classification(
    images=images,
    labels=labels,
    multi_label=False, # True if you have multi-label data
    fn_model=None, # a callable e.g. (function or transformers pipelines) that returns `str`
    fn_next_input=None, # a function that feeds gradio components actively with the next input
    csv_logger=False, # True if you want to log to a CSV
    dataset_name=None # "<my_hf_org>/<my_dataset>" if you want to log to the hub
)
interface.launch()

image-generation

from dataset_viber import AnnotatorInterFace

prompts = [
    "Anthony Bourdain laughing",
    "David Chang wearing a suit"
]
images = [
    "https://upload.wikimedia.org/wikipedia/commons/8/85/David_Chang_David_Shankbone_2010.jpg",
    "https://upload.wikimedia.org/wikipedia/commons/thumb/a/a5/Anthony_Bourdain_Peabody_2014b.jpg/440px-Anthony_Bourdain_Peabody_2014b.jpg",
]

interface = AnnotatorInterFace.for_image_generation(
    prompts=prompts,
    completions=images,
    fn_model=None, # a callable e.g. (function or transformers pipelines) that returns `str`
    fn_next_input=None, # a function that feeds gradio components actively with the next input
    csv_logger=False, # True if you want to log to a CSV
    dataset_name=None # "<my_hf_org>/<my_dataset>" if you want to log to the hub
)

interface.launch()

image-description

from dataset_viber import AnnotatorInterFace

images = [
    "https://upload.wikimedia.org/wikipedia/commons/thumb/a/a5/Anthony_Bourdain_Peabody_2014b.jpg/440px-Anthony_Bourdain_Peabody_2014b.jpg",
    "https://upload.wikimedia.org/wikipedia/commons/8/85/David_Chang_David_Shankbone_2010.jpg"
]
descriptions = ["Anthony Bourdain laughing", "David Chang wearing a suit"]

interface = AnnotatorInterFace.for_image_description(
    images=images,
    descriptions=descriptions, # optional to show initial descriptions
    fn_model=None, # a callable e.g. (function or transformers pipelines) that returns `str`
    fn_next_input=None, # a function that feeds gradio components actively with the next input
    csv_logger=False, # True if you want to log to a CSV
    dataset_name=None # "<my_hf_org>/<my_dataset>" if you want to log to the hub
)
interface.launch()

image-question-answering/visual-question-answering

from dataset_viber import AnnotatorInterFace

images = [
    "https://upload.wikimedia.org/wikipedia/commons/thumb/a/a5/Anthony_Bourdain_Peabody_2014b.jpg/440px-Anthony_Bourdain_Peabody_2014b.jpg",
    "https://upload.wikimedia.org/wikipedia/commons/8/85/David_Chang_David_Shankbone_2010.jpg"
]
questions = ["Who is this?", "What is he wearing?"]
answers = ["Anthony Bourdain", "a suit"]

interface = AnnotatorInterFace.for_image_question_answering(
    images=images,
    questions=questions, # optional to show initial questions
    answers=answers, # optional to show initial answers
    fn_model=None, # a callable e.g. (function or transformers pipelines) that returns `str`
    fn_next_input=None, # a function that feeds gradio components actively with the next input
    csv_logger=False, # True if you want to log to a CSV
    dataset_name=None # "<my_hf_org>/<my_dataset>" if you want to log to the hub
)
interface.launch()

image-generation-preference

from dataset_viber import AnnotatorInterFace

prompts = [
    "Anthony Bourdain laughing",
    "David Chang wearing a suit"
]

images_a = [
    "https://upload.wikimedia.org/wikipedia/commons/8/85/David_Chang_David_Shankbone_2010.jpg",
    "https://upload.wikimedia.org/wikipedia/commons/thumb/a/a5/Anthony_Bourdain_Peabody_2014b.jpg/440px-Anthony_Bourdain_Peabody_2014b.jpg",
]

images_b = [
    "https://upload.wikimedia.org/wikipedia/commons/thumb/a/a5/Anthony_Bourdain_Peabody_2014b.jpg/440px-Anthony_Bourdain_Peabody_2014b.jpg",
    "https://upload.wikimedia.org/wikipedia/commons/8/85/David_Chang_David_Shankbone_2010.jpg"
]

interface = AnnotatorInterFace.for_image_generation_preference(
    prompts=prompts,
    completions_a=images_a,
    completions_b=images_b,
    fn_model=None, # a callable e.g. (function or transformers pipelines) that returns `str`
    fn_next_input=None, # a function that feeds gradio components actively with the next input
    csv_logger=False, # True if you want to log to a CSV
    dataset_name=None # "<my_hf_org>/<my_dataset>" if you want to log to the hub
)
interface.launch()

BulkInterface

Built on top of the Dash, plotly-express, umap-learn, and Embedder to embed and understand your distribution and annotate your data.

https://github.com/user-attachments/assets/5e96c06d-e37f-45a0-9633-1a8e714d71ed

Hub dataset

text-visualization

from dataset_viber import BulkInterface
from datasets import load_dataset

ds = load_dataset("SetFit/ag_news", split="train[:2000]")

interface: BulkInterface = BulkInterface.for_text_visualization(
    ds.to_pandas()[["text", "label_text"]],
    content_column='text',
    label_column='label_text',
)
interface.launch()

text-classification

from dataset_viber import BulkInterface
from datasets import load_dataset

ds = load_dataset("SetFit/ag_news", split="train[:2000]")
df = ds.to_pandas()[["text", "label_text"]]

interface = BulkInterface.for_text_classification(
    dataframe=df,
    content_column='text',
    label_column='label_text',
    labels=df['label_text'].unique().tolist()
)
interface.launch()

chat-visualization

from dataset_viber.bulk import BulkInterface
from datasets import load_dataset

ds = load_dataset("argilla/distilabel-capybara-dpo-7k-binarized", split="train[:1000]")
df = ds.to_pandas()[["chosen"]]

interface = BulkInterface.for_chat_visualization(
    dataframe=df,
    chat_column='chosen',
)
interface.launch()

chat-classification

from dataset_viber.bulk import BulkInterface
from datasets import load_dataset

ds = load_dataset("argilla/distilabel-capybara-dpo-7k-binarized", split="train[:1000]")
df = ds.to_pandas()[["chosen"]]

interface = BulkInterface.for_chat_classification(
    dataframe=df,
    chat_column='chosen',
    labels=["math", "science", "history", "question seeking"],
)
interface.launch()

Embedder

Built on top of the onnx and optimum to efficiently embed data.

Embedder

from dataset_viber.embedder import Embedder

embedder = Embedder(model_id="sentence-transformers/all-MiniLM-L6-v2")
embedder.encode(["Anthony Bourdain was an amazing chef in New York."])

Utils

Shuffle inputs in the same order

When working with multiple inputs, you might want to shuffle them in the same order.

def shuffle_lists(*lists):
    if not lists:
        return []

    # Get the length of the first list
    length = len(lists[0])

    # Check if all lists have the same length
    if not all(len(lst) == length for lst in lists):
        raise ValueError("All input lists must have the same length")

    # Create a list of indices and shuffle it
    indices = list(range(length))
    random.shuffle(indices)

    # Reorder each list based on the shuffled indices
    return [
        [lst[i] for i in indices]
        for lst in lists
    ]

Random swap to randomize completions

When working with multiple completions, you might want to swap out the completions at the same index, where each completion index x is swapped with a random completion at the same index. This is useful for preference learning.

def swap_completions(*lists):
    # Assuming all lists are of the same length
    length = len(lists[0])

    # Check if all lists have the same length
    if not all(len(lst) == length for lst in lists):
        raise ValueError("All input lists must have the same length")

    # Convert the input lists (which are tuples) to a list of lists
    lists = [list(lst) for lst in lists]

    # Iterate over each index
    for i in range(length):
        # Get the elements at index i from all lists
        elements = [lst[i] for lst in lists]

        # Randomly shuffle the elements
        random.shuffle(elements)

        # Assign the shuffled elements back to the lists
        for j, lst in enumerate(lists):
            lst[i] = elements[j]

    return lists

Load remote image URLs from Hugging Face Hub

When working with images, you might want to load remote URLs from the Hugging Face Hub.

from datasets import Dataset, Image, load_dataset

dataset = load_dataset(
    "my_hf_org/my_image_dataset"
).cast_column("my_image_column", Image(decode=False))
dataset[0]["my_image_column"]
# {'bytes': None, 'path': 'path_to_image.jpg'}

Contribute and development setup

First, install PDM.

Then, install the environment, this will automatically create a .venv virtual env and install the dev environment.

pdm install

Lastly, run pre-commit for formatting on commit.

pre-commit install

Follow this guide on making first contributions.

References

Logo

Keyboard icons created by srip - Flaticon

Inspirations

Project details

Release history Release notifications | RSS feed

0.3.1

Sep 4, 2024

0.3

Sep 3, 2024

0.2.4

Aug 22, 2024

0.2.3

Aug 21, 2024

0.2.3rc1 pre-release

Aug 21, 2024

0.2.2

Aug 21, 2024

0.2.2rc3 pre-release

Aug 21, 2024

0.2.2rc2 pre-release

Aug 21, 2024

0.2.2rc1 pre-release

Aug 21, 2024

0.2.1

Aug 21, 2024

0.2.0

Aug 21, 2024

0.2.0rc8 pre-release

Aug 21, 2024

0.2.0rc7 pre-release

Aug 21, 2024

0.2.0rc6 pre-release

Aug 21, 2024

0.2.0rc5 pre-release

Aug 21, 2024

This version

0.2.0rc4 pre-release

Aug 21, 2024

0.2.0rc3 pre-release

Aug 16, 2024

0.2.0rc2 pre-release

Aug 16, 2024

0.2.0rc1 pre-release

Aug 16, 2024

0.2.0rc0 pre-release

Aug 16, 2024

0.1.2

Aug 14, 2024

0.1.1

Aug 14, 2024

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

dataset_viber-0.2.0rc4.tar.gz (36.3 kB view hashes)

Uploaded Aug 21, 2024 Source

Built Distribution

dataset_viber-0.2.0rc4-py3-none-any.whl (43.4 kB view hashes)

Uploaded Aug 21, 2024 Python 3

Hashes for dataset_viber-0.2.0rc4.tar.gz

Hashes for dataset_viber-0.2.0rc4.tar.gz
Algorithm	Hash digest
SHA256	`63a659756d3d0d03942e8592aad5173702223bea7accec2dea4e2238e7adde74`
MD5	`a40f4a34b013a6371151f4d2d3cca7fe`
BLAKE2b-256	`5ec1969450ae2cc2b6a7aeaf0c6c022d5266ebf9d3011657aa5bc944ec61040f`

Hashes for dataset_viber-0.2.0rc4-py3-none-any.whl

Hashes for dataset_viber-0.2.0rc4-py3-none-any.whl
Algorithm	Hash digest
SHA256	`5249740677f85e1c9c9b652ff26e569f39af22658517dc5e89f7d0ec666c986a`
MD5	`50a78c6bfd2e60b217f76f0cb98a0246`
BLAKE2b-256	`45dd7fccf317068c294a61bc25a4102f18fcc56c3e5c02651bfc55f3c6b16c9d`