app.py

import streamlit as st
from datasets import load_dataset_builder
from datasets import get_dataset_config_names
from os import listdir
from datasets import load_dataset, Dataset
from datasets_sql import query
import plotly.express as px
import numpy as np
import statistics

st.set_page_config(
    page_title="Evaluation Buddy",
    page_icon="./robot.png",
    layout="wide",
)

st.title("Hugging Face Evaluation Buddy")

top_datasets= ['glue', 'super_glue', 'wikitext', 'imdb', 'squad', 'squad_es', \
                'paws', 'librispeech_asr', 'wmt16', 'xnli', 'snli', 'ag_news', \
                'anli', 'amazon_polarity', 'squad_v2', 'conll2003', 'red_caps', \
                'common_voice', 'stsb_multi_mt', 'trec', 'tweet_eval', 'cosmos_qa',\
                'sick', 'xsum', 'wikiann', 'yelp_polarity', 'hellaswag', 'piqa', \
                'race', 'winogrande']

tasks= ['text-classification', 'question-answering-extractive', 'automatic-speech-recognition']
metrics= ['matthews_correlation', 'perplexity', 'meteor', 'code_eval', 'super_glue', 'rouge', 'mauve', 'cer', 'accuracy', 'recall', 'bleurt', 'sari', 'precision', 'mean_iou', 'squad', 'mahalanobis', 'chrf', 'mae', 'squad_v2', 'seqeval', 'cuad', 'wiki_split', 'google_bleu', 'competition_math', 'pearsonr', 'xtreme_s', 'comet', 'gleu', 'spearmanr', 'f1', 'frugalscore', 'bertscore', 'indic_glue', 'mse', 'xnli', 'ter', 'coval', 'wer', 'bleu', 'glue', 'sacrebleu']

with st.sidebar.expander("Datasets", expanded=True):
    dataset_name = st.selectbox(
        f"Choose a dataset to evaluate on:",
        sorted(top_datasets))
    configs = get_dataset_config_names(dataset_name)
    dataset_config = st.selectbox(
        f"Choose a configuration of your dataset:",
        configs)
    dataset_builder = load_dataset_builder(dataset_name, dataset_config)
    splits = [s for s in dataset_builder.info.splits]
    dataset_split = st.selectbox(
    f"Choose a dataset split:",
    splits)
    balanced_stdev = st.slider("Choose a standard deviation threshold for determining whether a dataset is balanced or not:", 0.00, 1.00, 0.20)



st.markdown("## Here is some information about your dataset:")

st.markdown("### Description")

st.markdown(dataset_builder.info.description)
st.markdown("For more information about this dataset, check out [its website](https://huggingface.co/datasets/"+dataset_name+")")

st.markdown("### Dataset-Specific Metrics")
if dataset_name in metrics:
    st.markdown("Great news! Your dataset has a dedicated metric for it! You can use it like this:")
    code = ''' from datasets import load_metric
 metric = load_metric('''+dataset+''', '''+config+''')'''
    st.code(code, language='python')
    dedicated_metric = True
else:
    st.markdown("Your dataset doesn't have a dedicated metric, but that's ok!")
    dedicated_metric = False

st.markdown("### Task-Specific Metrics")

try:
    task = dataset_builder.info.task_templates[0].task
    st.markdown("The task associated to it is: " + task)
    if task == 'automatic-speech-recognition':
        st.markdown('Automatic Speech Recognition has some dedicated metrics such as:')
        st.markdown('[Word Error Rate](https://huggingface.co/metrics/wer)')
        st.markdown('[Character Error Rate](https://huggingface.co/metrics/cer)')
    else:
        st.markdown("The task for your dataset doesn't have any dedicated metrics, but you can still use general ones!")
except:
    st.markdown("The task for your dataset doesn't have any dedicated metrics, but you can still use general ones!")


#print(dataset_builder.info.task_templates)
#print(dataset_builder.info.features)


#st.markdown("### General Metrics")



#dataset = load_dataset(dataset_name, dataset_config, dataset_split)
#print(dataset_name, dataset_config, dataset_split)

#print(labels.head())



try:
    num_classes = dataset_builder.info.features['label'].num_classes
    dataset = load_dataset(dataset_name, split=dataset_split)
    labels = query("SELECT COUNT(*) from dataset GROUP BY label").to_pandas()
    labels = labels.rename(columns={"count_star()": "count"})
    labels.index = dataset_builder.info.features['label'].names
    st.markdown("### Labelled  Metrics")
    st.markdown("Your dataset has "+ str(dataset_builder.info.features['label'].num_classes) + " labels : " + ', '.join(dataset_builder.info.features['label'].names))
    #TODO : figure out how to make a label plot
    st.plotly_chart(px.pie(labels, values = "count", names = labels.index, width=800, height=400))
    total = sum(c for c in labels['count'])
    proportion = [c/total for c in labels['count']]
    #proportion = [0.85, 0.15]
    stdev_dataset= statistics.stdev(proportion)
    if stdev_dataset <= balanced_stdev:
            st.markdown("Since your dataset is well-balanced, you can look at using:")
            st.markdown('[Accuracy](https://huggingface.co/metrics/accuracy)')
            accuracy_code = '''from datasets import load_metric
        metric = load_metric("accuracy")'''
            st.code(accuracy_code, language='python')

    else:
            st.markdown("Since your dataset is not well-balanced, you can look at using:")
            st.markdown('[F1 Score](https://huggingface.co/metrics/f1)')
            accuracy_code = '''from datasets import load_metric
        metric = load_metric("accuracy")'''
            st.code(accuracy_code, language='python')
            st.markdown('Since it takes into account both precision and recall, which works well to evaluate model performance on minority classes.')
except:
    st.markdown("### Unsupervised  Metrics")
    st.markdown("Since dataset doesn't have any labels, so the metrics that you can use for evaluation are:")
    st.markdown('[Perplexity](https://huggingface.co/metrics/perplexity)')
    perplexity_code = '''from datasets import load_metric
metric = load_metric("perplexity")'''
    st.code(perplexity_code, language='python')
    st.markdown('If you choose a model that was trained on **' + dataset_name + '** and use it to compute perplexity on text generated by your model, this can help determine how similar the two are.')