app.py

import streamlit as st
from datasets import load_dataset
import os 

HF_TOKEN = os.environ.get("HF_TOKEN", None)

st.set_page_config(page_title="SelfCheck", layout="wide")
st.title("SelfCheck scores")


@st.cache_data
def load_data(min_score=0.4, exclude_stories=True):
    ds = load_dataset("loubnabnl/comsop_450_samples_detailed", split="train", token=HF_TOKEN, num_proc=2)
    ds = ds.filter(lambda x: x["passage_score"] >= min_score)
    if exclude_stories:
        ds = ds.filter(lambda x: "story" not in x["format"])
    return ds

exclude_stories = st.checkbox("Exclude stories", False)
maximum_score = 0.7 if exclude_stories else 1.0
min_value = st.slider('Select minimum selfcheck score', 0.0, maximum_score, 0.1, key='min_score')

ds = load_data(min_score=min_value, exclude_stories=exclude_stories)
index = st.number_input(f'Found {len(ds)} samples, choose one', min_value=0, max_value=len(ds)-1, value=0, step=1)
min_score = st.number_input(f'Choose threshold for diplayed inconsistent sentences', min_value=0.2, max_value=1.0, value=0.4, step=0.1)

# Load data based on slider values and checkbox status
sample = ds[index]
st.markdown(f"**Passage Score:** {sample['passage_score']:.2f}, **seed data**: {sample['seed_data']}, **format**: {sample['format']}.")
st.markdown("---")
st.subheader("📕 Generated text")
st.markdown(sample['original_text'])

# get inconsistent sentences
st.subheader(f"🤔 Sentences with a high inconsistency score (> {min_score})")
sentences = sample["sentences_and_scores"]
sentences = [e for e in sentences if e["score"] > min_score]
sentences = sorted(sentences, key=lambda d: d['score'], reverse=True)

for i, s in enumerate(sentences):
    st.markdown(f"**Sentence {i}** with score {s['score']:.2f}:\n{s['sentence']}")
    st.markdown("---")