app.py

import streamlit as st
from transformers import pipeline

@st.cache_resource  # кэширование
def load_model():
    return pipeline("text-classification", model="voroninip/bert-paper-classifier-arxiv", top_k=None)

model = load_model()

def top_pct(preds, threshold=.95):
    preds = sorted(preds, key=lambda x: -x["score"])

    cum_score = 0
    for i, item in enumerate(preds):
        cum_score += item["score"]
        if cum_score >= threshold:
            break

    preds = preds[:(i+1)]

    return preds


def format_predictions(preds) -> str:
    """
    Prepare predictions and their scores for printing to the user
    """
    out = ""
    for i, item in enumerate(preds):
        out += f"{i+1}. {item['label']} (score {item['score']:.2f})\n"
    return out

st.markdown(
    """
    <style>
    .stApp {
        background: linear-gradient(to bottom right, #020420, #080f6b);
    }
    </style>
    """,
    unsafe_allow_html=True
)

st.markdown("""
    <div style='text-align: center;'>
        <img src='https://info.arxiv.org/brand/images/brand-logo-primary.jpg' alt='Centered Image' width='300'/>
    </div>
""", unsafe_allow_html=True)

st.markdown("""
    <h2 style='text-align: center; color: lime; font-family: Arial;'>
        🚀 arXiv paper categories predictor
    </h2>
""", unsafe_allow_html=True)

st.markdown("""
    <p style='
        color: white;
        font-size: 20px;
        font-family: "Courier New", monospace;
    '>
        Paste Title and Abstract of the paper and get most likely categories of the paper in the 
        <a href="https://arxiv.org/category_taxonomy" target="_blank" style="color: cyan; text-decoration: none;">
            arXiv taxonomy
        </a>
    </p>
""", unsafe_allow_html=True)

title = st.text_input("Title", value="")

abstract = st.text_input("Abstract", value="")

st.markdown("""
        <p style='
            color: white;
            font-size: 20px;
            font-family: "Courier New", monospace;
        '>
            Most likely categories of the paper:
        </p>
    """, unsafe_allow_html=True)

query = title + '\n' + abstract
if title or abstract:
    result = format_predictions(top_pct(model(query)[0]))
    st.write(result)