metadata
license: cc0-1.0
Note: Due to nature of toxic comments data and code contain explicit language.
Data is from kaggle, the Toxic Comment Classification Challenge
https://www.kaggle.com/competitions/jigsaw-toxic-comment-classification-challenge/data?select=train.csv.zip
A copy of the data exists on the data directory.
Trained over 20 epoch in a runpod
🤗 Running demo here:
https://huggingface.co/spaces/vluz/Tox
Code requires pandas, tensorflow, and streamlit. All can be installed via pip.
import os
import pickle
import streamlit as st
import tensorflow as tf
from tensorflow.keras.layers import TextVectorization
@st.cache_resource
def load_model():
model = tf.keras.models.load_model(os.path.join("model", "toxmodel.keras"))
return model
@st.cache_resource
def load_vectorizer():
from_disk = pickle.load(open(os.path.join("model", "vectorizer.pkl"), "rb"))
new_v = TextVectorization.from_config(from_disk['config'])
new_v.adapt(tf.data.Dataset.from_tensor_slices(["xyz"])) # fix for Keras bug
new_v.set_weights(from_disk['weights'])
return new_v
st.title("Toxic Comment Test")
st.divider()
model = load_model()
vectorizer = load_vectorizer()
default_prompt = "i love you man, but fuck you!"
input_text = st.text_area("Comment:", default_prompt, height=150).lower()
if st.button("Test"):
if not input_text:
st.write("⚠ Warning: Empty prompt.")
elif len(input_text) < 15:
st.write("⚠ Warning: Model is far less accurate with a small prompt.")
if input_text == default_prompt:
st.write("Expected results from default prompt are positive for 0 and 2")
with st.spinner("Testing..."):
inputv = vectorizer([input_text])
output = model.predict(inputv)
res = (output > 0.5)
st.write(["toxic","severe toxic","obscene","threat","insult","identity hate"], res)
st.write(output)
Put toxmodel.keras and vectorizer.pkl into the model dir.
Then do:
stramlit run toxtest.py
Expected result from default prompt is 0 and 2
Full code can be found here:
https://github.com/vluz/ToxTest/