import streamlit as st import pandas as pd from transformers import AutoTokenizer, AutoModelForSequenceClassification, pipeline # Function to load the pre-trained model def load_finetune_model(model_name): tokenizer = AutoTokenizer.from_pretrained(model_name) model = AutoModelForSequenceClassification.from_pretrained(model_name) return tokenizer, model def load_model(model_name): tokenizer = AutoTokenizer.from_pretrained(model_name) model = AutoModelForSequenceClassification.from_pretrained(model_name) sentiment_pipeline = pipeline("sentiment-analysis", tokenizer=tokenizer, model=model) return sentiment_pipeline # Streamlit app st.title("Multi-label Toxicity Detection App") st.write("Enter a text and select the fine-tuned model to get the toxicity analysis.") # Input text default_text = "You might be the most stupid person in the world." text = st.text_input("Enter your text:", value=default_text) category = {'LABEL_0': 'toxic', 'LABEL_1': 'severe_toxic', 'LABEL_2': 'obscene', 'LABEL_3': 'threat', 'LABEL_4': 'insult', 'LABEL_5': 'identity_hate'} # Model selection model_options = { "Olivernyu/finetuned_bert_base_uncased": { "description": "This model detects different types of toxicity like threats, obscenity, insults, and identity-based hate in text. The table is prepopulated with some data, the table will be displayed once you hit analyze.", }, "distilbert-base-uncased-finetuned-sst-2-english": { "labels": ["NEGATIVE", "POSITIVE"], "description": "This model classifies text into positive or negative sentiment. It is based on DistilBERT and fine-tuned on the Stanford Sentiment Treebank (SST-2) dataset.", }, "textattack/bert-base-uncased-SST-2": { "labels": ["LABEL_0", "LABEL_1"], "description": "This model classifies text into positive(LABEL_1) or negative(LABEL_0) sentiment. It is based on BERT and fine-tuned on the Stanford Sentiment Treebank (SST-2) dataset.", }, "cardiffnlp/twitter-roberta-base-sentiment": { "labels": ["LABEL_0", "LABEL_1", "LABEL_2"], "description": "This model classifies tweets into negative (LABEL_0), neutral(LABEL_1), or positive(LABEL_2) sentiment. It is based on RoBERTa and fine-tuned on a large dataset of tweets.", }, } selected_model = st.selectbox("Choose a fine-tuned model:", model_options) st.write("### Model Information") st.write(f"**Description:** {model_options[selected_model]['description']}") initial_table_df = pd.DataFrame(columns=["Text (portion)", "Toxicity class 1", "Class 1 probability", "Toxicity class 2", "Class 2 probability"]) initial_table_data = [{'Text (portion)': ["who's speaking? \n you goddamn cocksucker you know "], 'Toxicity class 1': ['obscene'], 'Class 1 probability': 0.7282997369766235, 'Toxicity class 2': ['toxic'], 'Class 2 probability': 0.2139672487974167}, {'Text (portion)': ['::Here is another source: Melissa Sue Halverson (2'], 'Toxicity class 1': ['toxic'], 'Class 1 probability': 0.24484945833683014, 'Toxicity class 2': ['obscene'], 'Class 2 probability': 0.1627064049243927}, {'Text (portion)': [', 8 November 2007 (UTC) \n\n All I can say is, havin'], 'Toxicity class 1': ['toxic'], 'Class 1 probability': 0.7277262806892395, 'Toxicity class 2': ['obscene'], 'Class 2 probability': 0.2502792477607727}, {'Text (portion)': ['::::I only see that at birth two persons are given'], 'Toxicity class 1': ['toxic'], 'Class 1 probability': 0.2711867094039917, 'Toxicity class 2': ['insult'], 'Class 2 probability': 0.15477754175662994}, {'Text (portion)': ["* There you have it: one man's Barnstar is another"], 'Toxicity class 1': ['toxic'], 'Class 1 probability': 0.5408656001091003, 'Toxicity class 2': ['insult'], 'Class 2 probability': 0.12563346326351166}, {'Text (portion)': ['" \n\n == Fact == \n\n Could just be abit of trivial f'], 'Toxicity class 1': ['toxic'], 'Class 1 probability': 0.35239243507385254, 'Toxicity class 2': ['obscene'], 'Class 2 probability': 0.1686778962612152}, {'Text (portion)': ['HE IS A GHAY ASS FUCKER@@!!'], 'Toxicity class 1': ['obscene'], 'Class 1 probability': 0.7819343209266663, 'Toxicity class 2': ['toxic'], 'Class 2 probability': 0.16951803863048553}, {'Text (portion)': ["I'VE SEEN YOUR CRIMES AGAINST CHILDREN AND I'M ASH"], 'Toxicity class 1': ['toxic'], 'Class 1 probability': 0.8491994738578796, 'Toxicity class 2': ['threat'], 'Class 2 probability': 0.04749392718076706}, {'Text (portion)': [':While with a lot of that essay says, general time'], 'Toxicity class 1': ['toxic'], 'Class 1 probability': 0.282654732465744, 'Toxicity class 2': ['obscene'], 'Class 2 probability': 0.15901680290699005}, {'Text (portion)': ['== Help == \n\n Please members of wiki, help me. My '], 'Toxicity class 1': ['toxic'], 'Class 1 probability': 0.3118911385536194, 'Toxicity class 2': ['obscene'], 'Class 2 probability': 0.16506287455558777}] for d in initial_table_data: initial_table_df = pd.concat([initial_table_df, pd.DataFrame(d)], ignore_index=True) # Load the model and perform toxicity analysis if "table" not in st.session_state: st.session_state['table'] = initial_table_df if st.button("Analyze"): if not text: st.write("Please enter a text.") else: with st.spinner("Analyzing toxicity..."): if selected_model == "Olivernyu/finetuned_bert_base_uncased": toxicity_detector = load_model(selected_model) outputs = toxicity_detector(text, top_k=2) category_names = ["toxic", "severe_toxic", "obscene", "threat", "insult", "identity_hate"] results = [] for item in outputs: results.append((category[item['label']], item['score'])) # Create a table with the input text (or a portion of it), the highest toxicity class, and its probability table_data = { "Text (portion)": [text[:50]], "Toxicity class 1": [results[0][0]], f"Class 1 probability": results[0][1], "Toxicity class 2": [results[1][0]], f"Class 2 probability": results[1][1] } # print("Before concatenation:") # print(table_df) # Concatenate the new data frame with the existing data frame st.session_state['table'] = pd.concat([pd.DataFrame(table_data), st.session_state['table']], ignore_index=True) # print("After concatenation:") # print(table_df) # Update the table with the new data frame st.table(st.session_state['table']) else: st.empty() sentiment_pipeline = load_model(selected_model) result = sentiment_pipeline(text) st.write(f"Sentiment: {result[0]['label']} (confidence: {result[0]['score']:.2f})") if result[0]['label'] in ['POSITIVE', 'LABEL_1'] and result[0]['score']> 0.9: st.balloons() elif result[0]['label'] in ['NEGATIVE', 'LABEL_0'] and result[0]['score']> 0.9: st.error("Hater detected.") else: st.write("Enter a text and click 'Analyze' to perform toxicity analysis.")