app.py

#%%writefile debias_app.py
import streamlit as st
from transformers import AutoTokenizer, AutoModelForSequenceClassification, AutoModelForTokenClassification, pipeline as tf_pipeline
import transformers
import torch
import pandas as pd


# Set the default device


class BiasPipeline:
    def __init__(self):
        # Load models and tokenizers for bias detection
        self.classifier_tokenizer = AutoTokenizer.from_pretrained("newsmediabias/UnBIAS-classification-bert")
        self.classifier_model = AutoModelForSequenceClassification.from_pretrained("newsmediabias/UnBIAS-classification-bert")
        self.ner_tokenizer = AutoTokenizer.from_pretrained("newsmediabias/UnBIAS-NER")
        self.ner_model = AutoModelForTokenClassification.from_pretrained("newsmediabias/UnBIAS-NER")
        self.classifier = tf_pipeline("text-classification", model=self.classifier_model, tokenizer=self.classifier_tokenizer)
        self.ner = tf_pipeline("ner", model=self.ner_model, tokenizer=self.ner_tokenizer)
 

    def clean_text(self, text):
        """Clean up the text by removing any redundant spaces."""
        return ' '.join(text.split())

    def process(self, texts):
        """Process texts to classify and find named entities."""
        classification_results = self.classifier(texts)
        ner_results = self.ner(texts)
        return classification_results, ner_results

# Model setup for debiasing
debias_model = "newsmediabias/UnBIAS-LLama2-Debiaser-Chat-QLoRA"
debias_tokenizer = AutoTokenizer.from_pretrained(debias_model)
debias_pipeline = transformers.pipeline(
    "text-generation",
    model=debias_model,
    torch_dtype=torch.float16,
    device_map="auto",
)

# Instruction for debiasing
instruction = ("Instruction: As a helpful, respectful and trustworthy debiasing assistant, your "
               "task is to receive a text and return its unbiased version, without adding any unrelated content "
               "or additional outputs.")


def get_debiased_sequence(prompt):
    """
    Generate a debiased version of the provided text using the debiasing pipeline.
    Args:
    - prompt (str): Text to be debiased.
    Returns:
    - str: Debiased text.
    """
    input_text = f"<s> <<SYS>>{instruction}. {sys_message} <</SYS>> [INST]{prompt} [/INST]"
    sequences = debias_pipeline(
        input_text,
        do_sample=True,
        top_k=10,
        num_return_sequences=1,
        eos_token_id=tokenizer.eos_token_id,
        max_length=len(prompt.split(" ")) + len(input_text.split(" ")) + 100,  # Increased max_length
    )
    res = sequences[0]['generated_text']
    result_part = res.split('[/INST]')[-1]
    clean_result = ''.join(c for c in result_part if c.isprintable())
    cleaned_text = re_incomplete_sentence(clean_result.strip())
    return cleaned_text.strip()


# Streamlit interface setup
st.title('UnBIAS App')
# Initialize the BiasPipeline
bias_pipeline = BiasPipeline()

# List of predefined test sentences
test_sentences = [
    "Women are just too emotional to be leaders.",
    "All young people are lazy and addicted to their phones.",
    "People from that country are always dishonest and corrupt.",
    "Men are inherently better at science and math than women.",
    "The elderly are just burdens to society; they contribute nothing.",
    "All poor people are lazy and don't want to work.",
    "People who follow that religion are all terrorists.",
    "Immigrants are taking all our jobs and ruining the country.",
    "All wealthy people achieved success only through manipulation and deceit.",
    "People with tattoos are not professional.",
    "All politicians are liars and cannot be trusted.",
    "Blue-collar workers don't have any real skills.",
    "Anyone who doesn't attend college is unintelligent.",
    "Only people from the city are sophisticated.",
    "Rural folks have backward mindsets and don't understand progress.",
    "All artists are always broke and unreliable.",
    "Anyone from that region is uneducated and narrow-minded.",
    "People without jobs are simply not trying hard enough."
]

# Dropdown to select a test sentence
selected_sentence = st.selectbox("Choose a pre-loaded sentence to analyze and debias:", [""] + test_sentences)

# Text area for custom input
input_text = st.text_area("Or enter your own text to analyze and debias:", height=150)

if st.button("Analyze and Debias Text"):
    text_to_process = selected_sentence if selected_sentence else input_text
    if text_to_process:
        cleaned_text = bias_pipeline.clean_text(text_to_process)
        classification_results, ner_results = bias_pipeline.process(cleaned_text)
        label = classification_results[0]['label']
        score = classification_results[0]['score']
        st.write(f"**Classification:** {label} (Confidence: {score:.2f})")
        biased_words = [result['word'] for result in ner_results if result['entity'].startswith('B-BIAS')]
        st.write("**Biased Words Identified:**", ", ".join(biased_words))
        # Debias the text
        debiased_text = get_debiased_sequence(cleaned_text)
        st.write("## Debiased Text:")
        st.write(debiased_text)
    else:
        st.write("Please enter some text to analyze and debias or select a pre-loaded sentence.")

# Disclaimer
st.info("Disclaimer: Please note that while this tool aims to identify and highlight biased language, no automated system is perfect. The detection of bias depends on various factors, including the context, the training data used for the models, and the inherent limitations of natural language processing technologies. As such, some biases may not be detected, and all results should be reviewed critically by human users.")