app.py

import arxiv
import gradio as gr
import pandas as pd
from apscheduler.schedulers.background import BackgroundScheduler
from cachetools import TTLCache, cached
from setfit import SetFitModel
from tqdm.auto import tqdm

CACHE_TIME = 60 * 60 * 12
MAX_RESULTS = 30_000


@cached(cache=TTLCache(maxsize=10, ttl=CACHE_TIME))
def get_arxiv_result():
    search = arxiv.Search(
        query="ti:dataset AND abs:machine learning",
        max_results=MAX_RESULTS,
        sort_by=arxiv.SortCriterion.SubmittedDate,
    )
    return [
        {
            "title": result.title,
            "abstract": result.summary,
            "url": result.entry_id,
            "category": result.primary_category,
            "updated": result.updated,
        }
        for result in tqdm(search.results(), total=MAX_RESULTS)
    ]


def load_model():
    return SetFitModel.from_pretrained("librarian-bots/is_new_dataset_teacher_model")


def format_row_for_model(row):
    return f"TITLE: {row['title']} \n\nABSTRACT: {row['abstract']}"


int2label = {0: "new_dataset", 1: "not_new_dataset"}


def get_predictions(data: list[dict], model=None, batch_size=32):
    if model is None:
        model = load_model()
    predictions = []
    for i in tqdm(range(0, len(data), batch_size)):
        batch = data[i : i + batch_size]
        text_inputs = [format_row_for_model(row) for row in batch]
        batch_predictions = model.predict_proba(text_inputs)
        for j, row in enumerate(batch):
            prediction = batch_predictions[j]
            row["prediction"] = int2label[int(prediction.argmax())]
            row["probability"] = float(prediction.max())
            predictions.append(row)
    return predictions


def create_markdown(row):
    title = row["title"]
    abstract = row["abstract"]
    arxiv_id = row["arxiv_id"]
    hub_paper_url = f"https://huggingface.co/papers/{arxiv_id}"
    updated = row["updated"]
    updated = updated.strftime("%Y-%m-%d")
    broad_category = row["broad_category"]
    category = row["category"]
    return f""" <h1> {title} </h1> updated: {updated} 
    | category: {broad_category}  | subcategory: {category} |
\n\n{abstract}
\n\n [Hugging Face Papers page]({hub_paper_url})
    """


@cached(cache=TTLCache(maxsize=100, ttl=CACHE_TIME))
def prepare_data():
    print("Downloading arxiv results...")
    arxiv_results = get_arxiv_result()
    print("loading model...")
    model = load_model()
    print("Making predictions...")
    predictions = get_predictions(arxiv_results, model=model)
    df = pd.DataFrame(predictions)
    df.loc[:, "arxiv_id"] = df["url"].str.extract(r"(\d+\.\d+)")
    df.loc[:, "broad_category"] = df["category"].str.split(".").str[0]
    df.loc[:, "markdown"] = df.apply(create_markdown, axis=1)
    return df


all_possible_arxiv_categories = prepare_data().category.unique().tolist()
broad_categories = prepare_data().broad_category.unique().tolist()


def create_markdown_summary(categories=broad_categories, all_categories=None):
    df = prepare_data()
    if categories is not None:
        df = df[df["broad_category"].isin(categories)]
    return "\n\n".join(df["markdown"].tolist())


scheduler = BackgroundScheduler()
scheduler.add_job(prepare_data, "cron", hour=3, minute=30)
scheduler.start()

with gr.Blocks() as demo:
    gr.Markdown("## New Datasets in Machine Learning")
    gr.Markdown(
        "This Space attempts to show new papers on arXiv that are *likely* to be papers"
        " introducing new datasets. \n\n"
    )
    broad_categories = gr.Dropdown(
        choices=broad_categories,
        label="Categories",
        multiselect=True,
        value=broad_categories,
    )
    results = gr.Markdown(create_markdown_summary())
    broad_categories.change(create_markdown_summary, broad_categories, results)

demo.launch()