app.py

import time
from itertools import count, islice
from functools import partial
from typing import Iterable, Iterator, TypeVar

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


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

MAX_TOTAL_NB_ITEMS = 100  # almost infinite, don't judge me (actually it's because gradio needs a fixed number of components)
MAX_NB_ITEMS_PER_GENERATION_CALL = 10

GENERATE_DATASET_NAMES_FOR_SEARCH_QUERY = (
        "A Machine Learning Practioner is looking for a dataset that matches '{search_query}'. "
        f"Generate a list of {MAX_NB_ITEMS_PER_GENERATION_CALL} 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:**."
)

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

landing_page_datasets_generated_text = """
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)
"""
default_output = landing_page_datasets_generated_text.strip().split("\n")
assert len(default_output) == MAX_NB_ITEMS_PER_GENERATION_CALL

css = """
a {
    color: var(--body-text-color);
}

.datasetButton {
    justify-content: start;
    justify-content: left;
}
.tags {
    font-size: var(--button-small-text-size);
    color: var(--body-text-color-subdued);
}
.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;
}
.insivibleButtonGroup {
    display: none;
}

@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;
}
"""


with gr.Blocks(css=css) as demo:
    generated_texts_state = gr.State((landing_page_datasets_generated_text,))
    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
        with gr.Row():
            with gr.Column(scale=4, min_width=0):
                pass
            with gr.Column(scale=10):
                button_groups: list[gr.Group] = []
                buttons: list[gr.Button] = []
                for i in range(MAX_TOTAL_NB_ITEMS):
                    if i < len(default_output):
                        line = default_output[i]
                        dataset_name, tags = line.split(".", 1)[1].strip(" )").split(" (", 1)
                        group_classes = "buttonsGroup"
                        dataset_name_classes = "topButton"
                        tags_classes = "bottomButton"
                    else:
                        dataset_name, tags = "⬜⬜⬜⬜⬜⬜", "░░░░, ░░░░, ░░░░"
                        group_classes = "buttonsGroup insivibleButtonGroup"
                        dataset_name_classes = "topButton linear-background"
                        tags_classes = "bottomButton linear-background"
                    with gr.Group(elem_classes=group_classes) as button_group:
                        button_groups.append(button_group)
                        buttons.append(gr.Button(dataset_name, elem_classes=dataset_name_classes))
                        buttons.append(gr.Button(tags, elem_classes=tags_classes))

                see_more = gr.Button("See more")  # TODO: dosable when reaching end of page
                gr.Markdown(f"_powered by [{model_id}](https://huggingface.co/{model_id})_")
            with gr.Column(scale=4, min_width=0):
                pass
        # more.click(search_more_datasets, inputs=[generated_texts, search_bar], outputs=[generated_texts] + 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")  # TODO: implement
                        back_button = gr.Button("< Back", size="sm")
                    with gr.Column(scale=4, min_width=0):
                        pass
            with gr.Column(scale=4, min_width=0):
                pass


    T = TypeVar("T")

    def batched(it: Iterable[T], n: int) -> Iterator[list[T]]:
        it = iter(it)
        while batch := list(islice(it, n)):
            yield batch


    def stream_reponse(msg: str, generated_texts: tuple[str] = (), max_tokens=500) -> Iterator[str]:
        messages = [
            {"role": "user", "content": msg}
        ] + [
            item
            for generated_text in generated_texts
            for item in [
                {"role": "assistant", "content": generated_text},
                {"role": "user", "content": "Can you generate more ?"},
            ]
        ]
        for _ in range(3):
            try:
                for message in client.chat_completion(
                    messages=messages,
                    max_tokens=max_tokens,
                    stream=True,
                    top_p=0.8,
                ):
                    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_line_by_line(search_query: str, generated_texts: tuple[str] = ()) -> Iterator[str]:
        search_query = search_query[:1000] if search_query.strip() else landing_page_query
        generated_text = ""
        current_line = ""
        for token in stream_reponse(
            GENERATE_DATASET_NAMES_FOR_SEARCH_QUERY.format(search_query=search_query),
            generated_texts=generated_texts,
        ):
            current_line += token
            if current_line.endswith("\n"):
                yield current_line
                generated_text += current_line
                current_line = ""
        yield current_line
        generated_text += current_line
        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 landing_page_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)

    search_datasets_inputs = search_bar
    search_datasets_outputs = button_groups + buttons + [generated_texts_state]

    def search_datasets(search_query):
        yield {generated_texts_state: []}
        yield {
            button_group: gr.Group(elem_classes="buttonsGroup insivibleButtonGroup")
            for button_group in button_groups[MAX_NB_ITEMS_PER_GENERATION_CALL:]
        }
        yield {
            k: v
            for dataset_name_button, tags_button in batched(buttons, 2)
            for k, v in {
                dataset_name_button: gr.Button("⬜⬜⬜⬜⬜⬜", elem_classes="topButton linear-background"),
                tags_button: gr.Button("░░░░, ░░░░, ░░░░", elem_classes="bottomButton linear-background")
            }.items()
        }
        current_item_idx = 0
        generated_text = ""
        for line in gen_datasets_line_by_line(search_query):
            if "I'm sorry" in line:
                raise gr.Error("Error: inappropriate content")
            if current_item_idx >= MAX_NB_ITEMS_PER_GENERATION_CALL:
                return
            if line.strip() and line.strip().split(".", 1)[0].isnumeric():
                try:
                    dataset_name, tags = line.strip().split(".", 1)[1].strip(" )").split(" (", 1)
                except ValueError:
                    dataset_name, tags = line.strip().split(".", 1)[1].strip(" )").split(" ", 1)
                dataset_name, tags = dataset_name.strip("()[]* "), tags.strip("()[]* ")
                generated_text += line
                yield {
                    buttons[2 * current_item_idx]: gr.Button(dataset_name, elem_classes="topButton"),
                    buttons[2 * current_item_idx + 1]: gr.Button(tags, elem_classes="bottomButton"),
                    generated_texts_state: (generated_text,),
                }
                current_item_idx += 1

    search_more_datasets_inputs = [search_bar, generated_texts_state]
    search_more_datasets_outputs = button_groups + buttons + [generated_texts_state]

    def search_more_datasets(search_query, generated_texts):
        current_item_idx = initial_item_idx = len(generated_texts) * MAX_NB_ITEMS_PER_GENERATION_CALL
        yield {
            button_group: gr.Group(elem_classes="buttonsGroup")
            for button_group in button_groups[len(generated_texts) * MAX_NB_ITEMS_PER_GENERATION_CALL:(len(generated_texts) + 1) * MAX_NB_ITEMS_PER_GENERATION_CALL]
        }
        generated_text = ""
        for line in gen_datasets_line_by_line(search_query, generated_texts=generated_texts):
            if "I'm sorry" in line:
                raise gr.Error("Error: inappropriate content")
            if current_item_idx - initial_item_idx >= MAX_NB_ITEMS_PER_GENERATION_CALL:
                return
            if line.strip() and line.strip().split(".", 1)[0].isnumeric():
                try:
                    dataset_name, tags = line.strip().split(".", 1)[1].strip(" )").split(" (", 1)
                except ValueError:
                    dataset_name, tags = line.strip().split(".", 1)[1].strip(" )").split(" ", 1) [0], ""
                dataset_name, tags = dataset_name.strip("()[]* "), tags.strip("()[]* ")
                generated_text += line
                yield {
                    buttons[2 * current_item_idx]: gr.Button(dataset_name, elem_classes="topButton"),
                    buttons[2 * current_item_idx + 1]: gr.Button(tags, elem_classes="bottomButton"),
                    generated_texts_state: (*generated_texts, generated_text),
                }
                current_item_idx += 1

    show_dataset_inputs = [search_bar, *buttons]
    show_dataset_outputs = [search_page, dataset_page, dataset_title, dataset_content]

    def show_dataset(search_query, *buttons_values, i):
        dataset_name, tags = buttons_values[2 * i : 2 * i + 2]
        yield {
            search_page: gr.Column(visible=False),
            dataset_page: gr.Column(visible=True),
            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_"
        }
        for generated_text in gen_dataset_content(search_query=search_query, dataset_name=dataset_name, tags=tags):
            yield {dataset_content: 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 ;)")


    search_bar.submit(search_datasets, inputs=search_datasets_inputs, outputs=search_datasets_outputs)
    search_button.click(search_datasets, inputs=search_datasets_inputs, outputs=search_datasets_outputs)
    for i, (dataset_name_button, tags_button) in enumerate(batched(buttons, 2)):
        dataset_name_button.click(partial(show_dataset, i=i), inputs=show_dataset_inputs, outputs=show_dataset_outputs)
        tags_button.click(partial(show_dataset, i=i), inputs=show_dataset_inputs, outputs=show_dataset_outputs)
    see_more.click(search_more_datasets, inputs=search_more_datasets_inputs, outputs=search_more_datasets_outputs)

    generate_full_dataset_button.click(generate_full_dataset)
    back_button.click(show_search_page, inputs=[], outputs=[search_page, dataset_page])

demo.launch()