app.py

import gradio as gr
import json
import re
import torch
from transformers import AutoTokenizer, AutoModelForSequenceClassification

model_name = 'yuntian-deng/ak-paper-selection-deberta'
max_length = 512
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)
model.eval()
if torch.cuda.is_available():
    model.cuda()

validation_results = json.load(open('validation_results.json'))
scores, thresholds, precisions, recalls = validation_results['scores'], validation_results['thresholds'], validation_results['precisions'], validation_results['recalls']


def get_threshold_precision(score_):
    for score, threshold, precision, recall in zip(scores, thresholds, precisions, recalls):
        if score_ < score:
            break
        prev_score, prev_threshold, prev_precision, prev_recall = score, threshold, precision, recall
    if prev_threshold == prev_score:
        prev_threshold = score_
    return prev_threshold, prev_precision, prev_recall

def extract_arxiv_id(input_text):
    if 'arxiv.org' in input_text:
        parsed_url = urlparse(input_text)
        query = parse_qs(parsed_url.query)
        path_parts = parsed_url.path.split('/')
        if 'id_list' in query:
            return query['id_list'][0]
        elif path_parts[-2] in ['abs', 'pdf']:
            return path_parts[-1].replace('.pdf', '')
    return input_text

def fetch_arxiv_data(arxiv_id):
    time.sleep(3)  # Comply with arXiv API terms of usage
    query_url = f'http://export.arxiv.org/api/query?id_list={arxiv_id}'
    response = feedparser.parse(query_url)
    if response.entries:
        entry = response.entries[0]
        title = entry.title
        authors = ', '.join(author.name for author in entry.authors)
        abstract = entry.summary
        return title, authors, abstract
    return "", "", ""

def update_fields(url_or_id):
    arxiv_id = extract_arxiv_id(url_or_id)
    title, authors, abstract = fetch_arxiv_data(arxiv_id)
    return title, authors, abstract

def normalize_spaces(text):
    return re.sub(r'\s+', ' ', text).strip()

def fill_template(title, authors, abstract):
    title = normalize_spaces(title.replace('\n', ' '))
    authors = ', '.join([author.strip() for author in authors.split(',')])
    abstract = normalize_spaces(abstract.replace('\n', ' '))
    text = f"""Title: {title}
Authors: {authors}
Abstract: {abstract}"""
    return text

@torch.no_grad()
def model_inference(title, authors, abstract):
    text = fill_template(title, authors, abstract)
    print (text)
    inputs = tokenizer([text], return_tensors="pt", truncation=True, max_length=max_length)
    if torch.cuda.is_available():
        inputs = {key: value.cuda() for key, value in inputs.items()}
    outputs = model(**inputs)
    logits = outputs.logits
    probs = logits.softmax(dim=-1).view(-1)
    score = probs[1].item()
    return score

def predict(title, authors, abstract):
    # Your model prediction logic here
    score = model_inference(title, authors, abstract)
    
    # Calculate precision for scores >= the predicted score
    #selected = [d for d in validation_data if d['score'] >= score]
    #true_positives = sum(1 for d in selected if d['label'] == 1)
    #precision = true_positives / len(selected) if selected else 0
    threshold, precision, recall = get_threshold_precision(score)
    
    result = f"Your score: {score:.2f}.\nFor papers with score>={threshold:.2f}, {precision * 100:.2f}% are selected by AK.\nFor papers selected by AK, {recall * 100:.2f}% have score>={threshold:.2f}"
    
    return score, result

example_title = "WildChat: 1M ChatGPT Interaction Logs in the Wild"
example_authors = "Wenting Zhao, Xiang Ren, Jack Hessel, Claire Cardie, Yejin Choi, Yuntian Deng"
example_abstract = "Chatbots such as GPT-4 and ChatGPT are now serving millions of users. Despite their widespread use, there remains a lack of public datasets showcasing how these tools are used by a population of users in practice. To bridge this gap, we offered free access to ChatGPT for online users in exchange for their affirmative, consensual opt-in to anonymously collect their chat transcripts and request headers. From this, we compiled WildChat, a corpus of 1 million user-ChatGPT conversations, which consists of over 2.5 million interaction turns. We compare WildChat with other popular user-chatbot interaction datasets, and find that our dataset offers the most diverse user prompts, contains the largest number of languages, and presents the richest variety of potentially toxic use-cases for researchers to study. In addition to timestamped chat transcripts, we enrich the dataset with demographic data, including state, country, and hashed IP addresses, alongside request headers. This augmentation allows for more detailed analysis of user behaviors across different geographical regions and temporal dimensions. Finally, because it captures a broad range of use cases, we demonstrate the dataset’s potential utility in fine-tuning instruction-following models. WildChat is released at https://wildchat.allen.ai under AI2 ImpACT Licenses."
    
iface = gr.Interface(
    fn=predict,
    inputs=[gr.Textbox(label="Paper Title", placeholder="Enter paper title", value=example_title), 
            gr.Textbox(label="Authors (separated by comma)", placeholder="Enter authors (separated by comma)", value=example_authors), 
            gr.TextArea(label="Abstract", placeholder="Enter abstract", value=example_abstract),
            gr.Textbox(label="[Optional] Autofill using arXiv URL/ID", placeholder="[Optional] Autofill using arXiv URL/ID", on_change=update_fields, change_elements=["title", "authors", "abstract"]),
],
    outputs=[gr.Textbox(label="Predicted Score"), gr.Textbox(label="Predicted Selection Probability")],
    title="Paper Selection Prediction",
    description="Predict if @_akhaliq will select your paper into Hugging Face papers. Enter the title, authors, and abstract of your paper, or enter an arXiv URL/ID.",
    live=False,
)

iface.queue(concurrency_limit=1, max_size=20).launch(enable_queue=True)