app.py

import torch
from transformers import RobertaForMaskedLM, RobertaTokenizer
import gradio as gr

model = RobertaForMaskedLM.from_pretrained('roberta-base')
tokenizer = RobertaTokenizer.from_pretrained('roberta-base')


def sentence_perplexity(sentence):
    tokenized_sentence = tokenizer(sentence, return_tensors='pt')
    input_ids = tokenized_sentence['input_ids']

    with torch.no_grad():
        output = model(input_ids)
    logits = output.logits

    probabilities = torch.softmax(logits, dim=-1)
    true_token_probabilities = torch.gather(probabilities, 2, input_ids.unsqueeze(-1)).squeeze(-1)
    log_probs = torch.log(true_token_probabilities)

    # Ignore the [CLS] and [SEP] tokens
    log_probs = log_probs[:, 1:-1]

    # Calculate the perplexity
    perplexity = torch.exp(-log_probs.mean()).item()

    return perplexity

def weird_score(sentence):
    perplexity = sentence_perplexity(sentence)

    # Normalize the weird score using the perplexity value
    weird_score = (perplexity - 1) / (perplexity + 1) * 100

    return f"Weird Score: {weird_score:.2f}%"

# Example usage
sentence = "This is a normal sentence."
print(weird_score(sentence))

sentence = "Giraffes are known to be fluent in six languages."
print(weird_score(sentence))


iface = gr.Interface(
    fn=weird_score,
    inputs=gr.inputs.Textbox(lines=2, placeholder="Enter a sentence..."),
    outputs="text",
    title="RoBERTa Weird Score Calculator",
    description="This app calculates the weird score percentage of a sentence using RoBERTa."
)
iface.launch()