app.py

import time
from functools import partial
from typing import Iterator

import gradio as gr
import requests.exceptions
from huggingface_hub import InferenceClient


model_id = "microsoft/Phi-3-mini-4k-instruct"
client = InferenceClient(model_id)

GENERATE_DATASET_NAMES_FOR_SEARCH_QUERY = (
        "A Machine Learning Practioner is looking for a dataset that matches '{search_query}'. "
        "Generate a list of 10 names of quality dataset that don't exist but sound plausible and would "
        "be helpful. Feel free to reuse words from the query '{search_query}' to name the datasets. "
        "Every dataset should be about '{search_query}' and have descriptive tags/keywords including the ML task name associated to the dataset (classification, regression, anomaly detection, etc.). Use the following format:\n1. DatasetName1 (tag1, tag2, tag3)\n1. DatasetName2 (tag1, tag2, tag3)"
)

GENERATE_DATASET_CONTENT_FOR_SEARCH_QUERY_AND_NAME_AND_TAGS = (
    "A ML practitioner is looking for a dataset CSV after the query '{search_query}'. "
    "Generate the first 5 rows of a plausible and quality CSV for the dataset '{dataset_name}'. "
    "You can get inspiration from related keywords '{tags}' but most importantly the dataset should correspond to the query '{search_query}'. "
    "Focus on quality text content and and use a 'label' or 'labels' column if it makes sense (invent labels, avoid reusing the keywords, be accurate while labelling texts). "
    "Reply using a short description of the dataset with title **Dataset Description:** followed by the CSV content in a code block and with title **CSV Content Preview:**."
)

default_query = "various datasets on many different subjects and topics, from classification to language modeling, from science to sport to finance to news"


def stream_reponse(msg: str, max_tokens=500) -> Iterator[str]:
    for _ in range(3):
        try:
            for message in client.chat_completion(
                messages=[{"role": "user", "content": msg}],
                max_tokens=max_tokens,
                stream=True,
            ):
                yield message.choices[0].delta.content
        except requests.exceptions.ConnectionError as e:
            print(e + "\n\nRetrying in 1sec")
            time.sleep(1)
            continue
        break


def gen_datasets(search_query: str) -> Iterator[str]:
    search_query = search_query[:1000] if search_query.strip() else default_query
    generated_text = ""
    for token in stream_reponse(GENERATE_DATASET_NAMES_FOR_SEARCH_QUERY.format(search_query=search_query)):
        generated_text += token
        if generated_text.endswith("\n"):
            yield generated_text.strip()
    yield generated_text.strip()
    print("-----\n\n" + generated_text)


def gen_dataset_content(search_query: str, dataset_name: str, tags: str) -> Iterator[str]:
    search_query = search_query[:1000] if search_query.strip() else default_query
    generated_text = ""
    for token in stream_reponse(GENERATE_DATASET_CONTENT_FOR_SEARCH_QUERY_AND_NAME_AND_TAGS.format(
        search_query=search_query,
        dataset_name=dataset_name,
        tags=tags,
    ), max_tokens=1500):
        generated_text += token
        yield generated_text
    print("-----\n\n" + generated_text)


NB_ITEMS_PER_PAGE = 10

default_output = """
1. NewsEventsPredict (classification, media, trend)
2. FinancialForecast (economy, stocks, regression)
3. HealthMonitor (science, real-time, anomaly detection)
4. SportsAnalysis (classification, performance, player tracking)
5. SciLiteracyTools (language modeling, science literacy, text classification)
6. RetailSalesAnalyzer (consumer behavior, sales trend, segmentation)
7. SocialSentimentEcho (social media, emotion analysis, clustering)
8. NewsEventTracker (classification, public awareness, topical clustering)
9. HealthVitalSigns (anomaly detection, biometrics, prediction)
10. GameStockPredict (classification, finance, sports contingency)
""".strip().split("\n")
assert len(default_output) == NB_ITEMS_PER_PAGE

css = """
.datasetButton {
    justify-content: start;
    justify-content: left;
}
.tags {
    font-size: var(--button-small-text-size);
    color: var(--body-text-color-subdued);
}
a {
    color: var(--body-text-color);
}
.topButton {
    justify-content: start;
    justify-content: left;
    text-align: left;
    background: transparent;
    box-shadow: none;
    padding-bottom: 0;
}
.topButton::before {
    content: url("data:image/svg+xml,%3Csvg style='color: rgb(209 213 219)' xmlns='http://www.w3.org/2000/svg' xmlns:xlink='http://www.w3.org/1999/xlink' aria-hidden='true' focusable='false' role='img' width='1em' height='1em' preserveAspectRatio='xMidYMid meet' viewBox='0 0 25 25'%3E%3Cellipse cx='12.5' cy='5' fill='currentColor' fill-opacity='0.25' rx='7.5' ry='2'%3E%3C/ellipse%3E%3Cpath d='M12.5 15C16.6421 15 20 14.1046 20 13V20C20 21.1046 16.6421 22 12.5 22C8.35786 22 5 21.1046 5 20V13C5 14.1046 8.35786 15 12.5 15Z' fill='currentColor' opacity='0.5'%3E%3C/path%3E%3Cpath d='M12.5 7C16.6421 7 20 6.10457 20 5V11.5C20 12.6046 16.6421 13.5 12.5 13.5C8.35786 13.5 5 12.6046 5 11.5V5C5 6.10457 8.35786 7 12.5 7Z' fill='currentColor' opacity='0.5'%3E%3C/path%3E%3Cpath d='M5.23628 12C5.08204 12.1598 5 12.8273 5 13C5 14.1046 8.35786 15 12.5 15C16.6421 15 20 14.1046 20 13C20 12.8273 19.918 12.1598 19.7637 12C18.9311 12.8626 15.9947 13.5 12.5 13.5C9.0053 13.5 6.06886 12.8626 5.23628 12Z' fill='currentColor'%3E%3C/path%3E%3C/svg%3E");
    margin-right: .25rem;
    margin-left: -.125rem;
    margin-top: .25rem;
}
.bottomButton {
    justify-content: start;
    justify-content: left;
    text-align: left;
    background: transparent;
    box-shadow: none;
    font-size: var(--button-small-text-size);
    color: var(--body-text-color-subdued);
    padding-top: 0;
    align-items: baseline;
}
.bottomButton::before {
    content: 'tags:';
    margin-right: .25rem;
}
.buttonsGroup {
    background: transparent;
}
.buttonsGroup:hover {
    background: var(--input-background-fill);
}
.buttonsGroup div {
    background: transparent;
}
@keyframes placeHolderShimmer{
    0%{
        background-position: -468px 0
    }
    100%{
        background-position: 468px 0
    }
}
.linear-background {
    animation-duration: 1s;
    animation-fill-mode: forwards;
    animation-iteration-count: infinite;
    animation-name: placeHolderShimmer;
    animation-timing-function: linear;
    background-image: linear-gradient(to right, var(--body-text-color-subdued) 8%, #dddddd11 18%, var(--body-text-color-subdued) 33%);
    background-size: 1000px 104px;
    color: transparent;
    background-clip: text;
}
"""

def search_datasets(search_query):
    output_values = [
        gr.Button("⬜⬜⬜⬜⬜⬜", elem_classes="topButton linear-background"),
        gr.Button("░░░░, ░░░░, ░░░░", elem_classes="bottomButton linear-background")
    ] * NB_ITEMS_PER_PAGE
    for generated_text in gen_datasets(search_query):
        if "I'm sorry" in generated_text:
            raise gr.Error("Error: inappropriate content")
        lines = [line for line in generated_text.split("\n") if line and line.split(".", 1)[0].isnumeric()][:NB_ITEMS_PER_PAGE]
        for i, line in enumerate(lines):
            dataset_name, tags = line.split(".", 1)[1].strip(" )").split(" (", 1)
            output_values[2 * i] = gr.Button(dataset_name, elem_classes="topButton")
            output_values[2 * i + 1] = gr.Button(tags, elem_classes="bottomButton")
        yield output_values


def show_dataset(search_query, *buttons_values, i):
    dataset_name, tags = buttons_values[2 * i : 2 * i + 2]
    dataset_title = f"# {dataset_name}\n\n tags: {tags}\n\n _Note: This is an AI-generated dataset so its content may be inaccurate or false_"
    yield gr.Column(visible=False), gr.Column(visible=True), dataset_title, ""
    for generated_text in gen_dataset_content(search_query=search_query, dataset_name=dataset_name, tags=tags):
        yield gr.Column(), gr.Column(), dataset_title, generated_text


def show_search_page():
    return gr.Column(visible=True), gr.Column(visible=False)


def generate_full_dataset():
    raise gr.Error("Not implemented yet sorry ! Give me some feedbacks in the Community tab in the meantime ;)")


with gr.Blocks(css=css) as demo:
    with gr.Row():
        with gr.Column(scale=4, min_width=0):
            pass
        with gr.Column(scale=10):
            gr.Markdown(
                "# 🤗 Infinite Dataset Hub ♾️\n\n"
                "An endless catalog of datasets, created just for you.\n\n"
            )
        with gr.Column(scale=4, min_width=0):
            pass
    with gr.Column() as search_page:
        with gr.Row():
            with gr.Column(scale=4, min_width=0):
                pass
            with gr.Column(scale=9):
                search_bar = gr.Textbox(max_lines=1, placeholder="Search datasets, get infinite results", show_label=False, container=False)
            with gr.Column(min_width=64):
                search_button = gr.Button("🔍", variant="primary")
            with gr.Column(scale=4, min_width=0):
                pass
        inputs = [search_bar]
        show_dataset_outputs = [search_page]
        with gr.Row():
            with gr.Column(scale=4, min_width=0):
                pass
            with gr.Column(scale=10):
                buttons = []
                for i in range(10):
                    line = default_output[i]
                    dataset_name, tags = line.split(".", 1)[1].strip(" )").split(" (", 1)
                    with gr.Group(elem_classes="buttonsGroup"):
                        top = gr.Button(dataset_name, elem_classes="topButton")
                        bottom = gr.Button(tags, elem_classes="bottomButton")
                        buttons += [top, bottom]
                        top.click(partial(show_dataset, i=i), inputs=inputs, outputs=show_dataset_outputs)
                        bottom.click(partial(show_dataset, i=i), inputs=inputs, outputs=show_dataset_outputs)
                inputs += buttons
                gr.Markdown(f"_powered by [{model_id}](https://huggingface.co/{model_id})_")
            with gr.Column(scale=4, min_width=0):
                pass
        search_bar.submit(search_datasets, inputs=search_bar, outputs=buttons)
        search_button.click(search_datasets, inputs=search_bar, outputs=buttons)
    with gr.Column(visible=False) as dataset_page:
        with gr.Row():
            with gr.Column(scale=4, min_width=0):
                pass
            with gr.Column(scale=10):
                dataset_title = gr.Markdown()
                dataset_content = gr.Markdown()
                with gr.Row():
                    with gr.Column(scale=4, min_width=0):
                        pass
                    with gr.Column():
                        generate_full_dataset_button = gr.Button("Generate Full Dataset", variant="primary")
                        generate_full_dataset_button.click(generate_full_dataset)
                        back_button = gr.Button("< Back", size="sm")
                        back_button.click(show_search_page, inputs=[], outputs=[search_page, dataset_page])
                    with gr.Column(scale=4, min_width=0):
                        pass
            with gr.Column(scale=4, min_width=0):
                pass
        show_dataset_outputs += [dataset_page, dataset_title, dataset_content]
demo.launch()