app.py

import streamlit as st
import datasets
import os
import json
from transformers import AutoTokenizer
import ast
import re

version = st.sidebar.selectbox("Choose a version", ["init","local_dedup"])
if version == "init":
    CACHE_DIR = "cache_ds/" #Use this to build the dataset
else:
    CACHE_DIR = "local_dedup/"
contribution_json = "contributors.json"

contribution_dict = json.load(open(contribution_json,"r"))
IGNORE_LIST = ["Bible","Tanzil","GNOME"]

splits = [split for split in os.listdir(CACHE_DIR) if split not in IGNORE_LIST]

cached_ds = os.listdir(CACHE_DIR)
tokenizer = AutoTokenizer.from_pretrained('EleutherAI/gpt-neox-20b')


def load_page(split):
    with st.spinner('Downloading and buidling dataset...'):
        if split not in cached_ds:
            ds = datasets.load_dataset('CarperAI/pile-v2-small-filtered',"train", data_files="data/"+split+"/data.json")
        else:
            ds = datasets.load_from_disk(CACHE_DIR+split)
    print("Sucessfully loaded "+split)
    st.title("Dataset Explorer")
    st.write(f"# {split}")
    if split in contribution_dict:
        st.caption(f"Contributors: {','.join(contribution_dict[split])}")
    else:
        st.caption(f"Needs to be updated....")
    with st.form("dataset_form"):
        index = st.slider('Select a row', 0, len(ds)-1, 0)
        if st.form_submit_button("Load"):
            st.write(f"Row {index}")
            data = ds[index]
            content = data["text"]
            meta = data["meta"]
            with st.expander("Render Content"):
                st.write(content)
            with st.expander("Raw Content"):
                st.text(content)
            with st.expander("Metadata and Metrics"):    
                st.write("### Meta:")
                try:
                    st.write(ast.literal_eval(meta))
                except:
                    st.write(meta)
            # Tokenizer-related count
                tokenized =  tokenizer(content, return_length=True)['length'][0]
                token_count_metric = st.metric("Token Count(compared to 2048)",value=tokenized,delta=4096-tokenized)
            #Word related count
                split_words = re.findall(r'\w+', content)
                word_count_metric = st.metric("Word Count",value=len(split_words))
                


demo_name = st.sidebar.selectbox("Choose a demo", splits)
load_page(demo_name)