app.py

import requests
from collections import Counter
from requests.adapters import HTTPAdapter, Retry
import multiprocessing
import os
import time
import logging

import gradio as gr
import pandas as pd
import polars as pl
import matplotlib.pyplot as plt
import spaces
from gradio_huggingfacehub_search import HuggingfaceHubSearch
from huggingface_hub import PyTorchModelHubMixin
import torch
from torch import nn
from transformers import AutoModel, AutoTokenizer, AutoConfig
from tqdm import tqdm


logging.basicConfig(level=logging.INFO, format="%(asctime)s - %(name)s - %(levelname)s - %(message)s")


session = requests.Session()
retries = Retry(total=5, backoff_factor=1, status_forcelist=[502, 503, 504])
session.mount('http://', HTTPAdapter(max_retries=retries))


class QualityModel(nn.Module, PyTorchModelHubMixin):
    def __init__(self, config):
        super(QualityModel, self).__init__()
        self.model = AutoModel.from_pretrained(config["base_model"])
        self.dropout = nn.Dropout(config["fc_dropout"])
        self.fc = nn.Linear(self.model.config.hidden_size, len(config["id2label"]))

    def forward(self, input_ids, attention_mask):
        features = self.model(
            input_ids=input_ids, attention_mask=attention_mask
        ).last_hidden_state
        dropped = self.dropout(features)
        outputs = self.fc(dropped)
        return torch.softmax(outputs[:, 0, :], dim=1)

device = "cuda" if torch.cuda.is_available() else "cpu"
config = AutoConfig.from_pretrained("nvidia/quality-classifier-deberta")
tokenizer = AutoTokenizer.from_pretrained("nvidia/quality-classifier-deberta")
model = QualityModel.from_pretrained("nvidia/quality-classifier-deberta").to(device)
# model = torch.compile(model)
model.eval()


@spaces.GPU
def predict(texts: list[str]):
    inputs = tokenizer(
        texts, return_tensors="pt", padding="longest", truncation=True
    ).to(device)
    outputs = model(inputs["input_ids"], inputs["attention_mask"])
    predicted_classes = torch.argmax(outputs, dim=1)
    predicted_domains = [
        config.id2label[class_idx.item()] for class_idx in predicted_classes.cpu().numpy()
    ]
    return predicted_domains


def plot_and_df(texts, preds):
    texts_df = pd.DataFrame({"quality": preds, "text": texts})
    counts = Counter(preds)
    counts_df = pd.DataFrame(
        {
            "quality": ["Low", "Medium", "High"],
            "count": [counts.get("Low", 0), counts.get("Medium", 0), counts.get("High", 0)]
        }
    )
    # counts.reset_index(inplace=True)
    return (
            gr.BarPlot(counts_df, x="quality", y="count", sort=None),
            texts_df[texts_df["quality"] == "Low"][["text"]][:20],
            texts_df[texts_df["quality"] == "Medium"][["text"]][:20],
            texts_df[texts_df["quality"] == "High"][["text"]][:20],
        )


@spaces.GPU
def run_quality_check(dataset, config, split, column, batch_size, num_examples):
    logging.info(f"Fetching data for {dataset=} {config=} {split=} {column=}")
    try:
        data = pl.read_parquet(f"hf://datasets/{dataset}@~parquet/{config}/{split}/0000.parquet", columns=[column])
    except pl.exceptions.ComputeError:
        try:
            data = pl.read_parquet(f"hf://datasets/{dataset}@~parquet/{config}/partial-{split}/0000.parquet", columns=[column])
        except pl.exceptions.ComputeError:
            try:
                data = pl.read_parquet(f"hf://datasets/{dataset}@~parquet/{config}/{split}-part0/0000.parquet", columns=[column])
            except Exception as error:
                yield f"❌ {error}", gr.BarPlot(), pd.DataFrame(), pd.DataFrame(), pd.DataFrame(), pd.DataFrame(),
                return
    logging.info("Data fetched.")

    data_sample = data.sample(num_examples, seed=16) if data.shape[0] > num_examples else data
    texts = [text[:10000] for text in data_sample[column].to_list()]
    predictions, texts_processed = [], []
    num_examples = min(len(texts), num_examples)
    for i in range(0, num_examples, batch_size):
        batch_texts = texts[i:i+batch_size]
        batch_predictions = predict(batch_texts)
        predictions.extend(batch_predictions)
        texts_processed.extend(batch_texts)
        yield {"check in progress...": i / num_examples}, *plot_and_df(texts_processed, predictions), pd.DataFrame()

    # with multiprocessing.Pool(processes=8) as pool:
    #     props = pool.map(proportion_non_ascii, texts)
    #
    # # non_ascii_df = pd.DataFrame.from_dict({"prop_non_ascii": props, "text": texts})
    # plt.hist(props, bins=20, range=(0., 1.))
    # plt.title('Histogram of proportion of non-ASCII characters')
    # plt.xlabel('Proportion of non-ASCII characters')
    # plt.ylabel('Number of texts')

    yield {"finished": 1.}, *plot_and_df(texts_processed, predictions), data_sample


PERSPECTIVE_API_KEY = os.environ.get("PERSPECTIVE_API_KEY")
PERSPECTIVE_URL = f"https://commentanalyzer.googleapis.com/v1alpha1/comments:analyze?key={PERSPECTIVE_API_KEY}"
REQUESTED_ATTRIBUTES = {"TOXICITY": {}, "SEVERE_TOXICITY": {},
                        "IDENTITY_ATTACK": {}, "INSULT": {}, "PROFANITY": {},
                        "THREAT": {}}
ATT_SCORE = "attributeScores"
SUM_SCORE = "summaryScore"


def plot_toxicity(scores):
    fig, axs = plt.subplots(2, 3)#, figsize=(10, 6))
    for x, y, score_name in zip([0,0,0,1,1,1], [0,1,2,0,1,2], scores):
        axs[x,y].hist(scores[score_name], bins=20, range=(0., 1.))
        axs[x,y].set_xlabel(score_name)
    fig.supylabel("Number of texts")
    fig.suptitle("Histogram of toxicity scores")
    fig.tight_layout()

    return fig

def call_perspective_api(texts_df, column_name, dataset, config, split):#, full_check=False):
    headers = {
        "content-type": "application/json",
    }
    req_att_scores = {attr: [] for attr in REQUESTED_ATTRIBUTES}

    # fetch data if it doesn't exist yet
    if texts_df.values.tolist() == [['', '', '']]:
        logging.info(f"Fetching data for {dataset=} {config=} {split=} {column_name=}")
        try:
            texts_df = pl.read_parquet(f"hf://datasets/{dataset}@~parquet/{config}/{split}/0000.parquet", columns=[column_name])
        except pl.exceptions.ComputeError:
            try:
                texts_df = pl.read_parquet(f"hf://datasets/{dataset}@~parquet/{config}/partial-{split}/0000.parquet", columns=[column_name])
            except pl.exceptions.ComputeError:
                try:
                    texts_df = pl.read_parquet(f"hf://datasets/{dataset}@~parquet/{config}/{split}-part0/0000.parquet", columns=[column_name])
                except Exception as error:
                    yield f"❌ {error}", plt.gcf(), pd.DataFrame(),
                    return
        logging.info("Data fetched.")
        texts_df = texts_df.to_pandas()

    # texts = texts_df.sample(100, seed=16)[column_name].values if not full_check else texts_df[column_name].values
    texts = texts_df.sample(100, random_state=16)[column_name].values if texts_df.shape[0] > 100 else texts_df[column_name].values

    n_samples = len(texts)
    for i, text in tqdm(enumerate(texts), desc="scanning with perspective"):
        data = {
            "comment": {"text": text},
            "languages": ["en"],
            "requestedAttributes": REQUESTED_ATTRIBUTES
        }
        time.sleep(1)
        try:
            req_response = requests.post(PERSPECTIVE_URL, json=data, headers=headers)
        except Exception as e:
            print(e)
            return req_att_scores

        if req_response.ok:
            response = req_response.json()
            if ATT_SCORE in response:
                for req_att in REQUESTED_ATTRIBUTES:
                    if req_att in response[ATT_SCORE]:
                        att_score = response[ATT_SCORE][req_att][SUM_SCORE]["value"]
                        req_att_scores[req_att].append(att_score)
                    else:
                        req_att_scores[req_att].append(0)
            else:
                raise ValueError(req_response)
        else:
            try:
                req_response.raise_for_status()
            except Exception as e:
                logging.info(e)
                return req_att_scores
        if i % 10 == 0:
            plot_toxicity(req_att_scores)
            print(len(texts[:i]), len(req_att_scores["TOXICITY"]))
            yield {"toxicity check in progress...": i / n_samples}, plt.gcf(), pd.DataFrame.from_dict({column_name: texts[:i+1], **req_att_scores})

    plot_toxicity(req_att_scores)
    yield {"toxicity check finished.": 1.}, plt.gcf(), pd.DataFrame.from_dict({column_name: texts, **req_att_scores})


def proportion_non_ascii(s):
    """
    Compute the proportion of non-ASCII characters in a string.

    Parameters:
    s (str): The input string.

    Returns:
    float: The proportion of non-ASCII characters in the string.
    """
    non_ascii_count = sum(1 for c in s if ord(c) > 127)
    total_chars = len(s)
    return non_ascii_count / total_chars if total_chars > 0 else 0.0


def non_ascii_check(texts_df, column_name):
    texts = texts_df[column_name].to_list()
    with multiprocessing.Pool(processes=8) as pool:
        props = pool.map(proportion_non_ascii, texts)

    # non_ascii_df = pd.DataFrame.from_dict({"prop_non_ascii": props, "text": texts})
    plt.hist(props, bins=20, range=(0., 1.))
    plt.title('Histogram of proportion of non-ASCII characters')
    plt.xlabel('Proportion of non-ASCII characters')
    plt.ylabel('Number of texts')

    return plt.gcf()


with gr.Blocks() as demo:
    gr.Markdown(
        """
        # 💫 Dataset Quality Checker 💫
        Use [nvidia/quality-classifier-deberta](https://huggingface.co/nvidia/quality-classifier-deberta) on any text dataset on the Hub.
        ## Select dataset and text column
        """
    )
    with gr.Row():
        with gr.Column(scale=3):
            dataset_name = HuggingfaceHubSearch(
                    label="Hub Dataset ID",
                    placeholder="Search for dataset id on Huggingface",
                    search_type="dataset",
                )
        subset_dropdown = gr.Dropdown(label="Subset", visible=False)
        split_dropdown = gr.Dropdown(label="Split", visible=False)

    # config_name = "default"  # TODO: user input
    with gr.Accordion("Dataset preview", open=False):
        @gr.render(inputs=[dataset_name, subset_dropdown, split_dropdown])
        def embed(name, subset, split):
            html_code = f"""
            <iframe
              src="https://huggingface.co/datasets/{name}/embed/viewer/{subset}/{split}"
              frameborder="0"
              width="100%"
              height="600px"
            ></iframe>
                """
            return gr.HTML(value=html_code)

    text_column_dropdown = gr.Dropdown(label="Text column name", info="Text colum name to check (only non-nested texts are supported)")

    def _resolve_dataset_selection(dataset: str, default_subset: str, default_split: str):
        if "/" not in dataset.strip().strip("/"):
            return {
                subset_dropdown: gr.Dropdown(visible=False),
                split_dropdown: gr.Dropdown(visible=False),
                text_column_dropdown: gr.Dropdown(info="Text colum name to check (only non-nested texts are supported)"),
            }
        info_resp = session.get(f"https://datasets-server.huggingface.co/info?dataset={dataset}", timeout=3).json()
        if "error" in info_resp:
            return {
                subset_dropdown: gr.Dropdown(visible=False),
                split_dropdown: gr.Dropdown(visible=False),
                text_column_dropdown: gr.Dropdown(label="Text column name", info="Text colum name to check (only non-nested texts are supported)")
            }
        subsets: list[str] = list(info_resp["dataset_info"])
        subset = default_subset if default_subset in subsets else subsets[0]
        splits: list[str] = info_resp["dataset_info"][subset]["splits"]
        split = default_split if default_split in splits else splits[0]
        features = info_resp["dataset_info"][subset]["features"]
        text_features = [feature_name for feature_name, feature in features.items() if isinstance(feature, dict) and feature.get("dtype") == "string"] # and feature.get("_type") == "Value"]
        return {
            subset_dropdown: gr.Dropdown(value=subset, choices=subsets, visible=len(subsets) > 1),
            split_dropdown: gr.Dropdown(value=split, choices=splits, visible=len(splits) > 1),
            text_column_dropdown: gr.Dropdown(choices=text_features, label="Text column name", info="Text colum name to check (only non-nested texts are supported)"),
        }

    @dataset_name.change(inputs=[dataset_name], outputs=[subset_dropdown, split_dropdown, text_column_dropdown])
    def show_input_from_subset_dropdown(dataset: str) -> dict:
        return _resolve_dataset_selection(dataset, default_subset="default", default_split="train")

    @subset_dropdown.change(inputs=[dataset_name, subset_dropdown], outputs=[subset_dropdown, split_dropdown, text_column_dropdown])
    def show_input_from_subset_dropdown(dataset: str, subset: str) -> dict:
        return _resolve_dataset_selection(dataset, default_subset=subset, default_split="train")

    @split_dropdown.change(inputs=[dataset_name, subset_dropdown, split_dropdown], outputs=[subset_dropdown, split_dropdown, text_column_dropdown])
    def show_input_from_split_dropdown(dataset: str, subset: str, split: str) -> dict:
        return _resolve_dataset_selection(dataset, default_subset=subset, default_split=split)

    gr.Markdown("## Run nvidia quality classifier")
    batch_size = gr.Slider(0, 64, 32, step=4, label="Inference batch size", info="(set this to smaller value if this space crashes.)")
    num_examples = gr.Slider(0, 1000, 500, step=10, label="Number of examples", info="Number of random examples to run quality classifier on")
    gr_check_btn = gr.Button("Check Dataset")
    progress_bar = gr.Label(show_label=False)
    plot = gr.BarPlot()


    with gr.Accordion("Explore some individual examples for each class", open=False):
        gr.Markdown("### Low")
        df_low = gr.DataFrame()
        gr.Markdown("### Medium")
        df_medium = gr.DataFrame()
        gr.Markdown("### High")
        df_high = gr.DataFrame()

    texts_df = gr.DataFrame(visible=False)

    gr.Examples(
        [
            ["HuggingFaceFW/fineweb-edu", "default", "train", "text", 16, 100],
            ["fka/awesome-chatgpt-prompts", "default", "train", "prompt", 64, 200],
            # ["Anthropic/hh-rlhf"],
            # ["OpenAssistant/oasst1"],
        ],
        [dataset_name, subset_dropdown, split_dropdown, text_column_dropdown, batch_size, num_examples],
        [progress_bar, plot, df_low, df_medium, df_high, texts_df],
        fn=run_quality_check,
        run_on_click=False,
        cache_examples="lazy",
    )

    gr_check_btn.click(
        run_quality_check,
        inputs=[dataset_name, subset_dropdown, split_dropdown, text_column_dropdown, batch_size, num_examples],
        outputs=[progress_bar, plot, df_low, df_medium, df_high, texts_df]
    )

    # gr.Markdown("""## Compute text quality measures
    #             * proportion of non-ascii characters
    #             * #TODO""")
    # gr_ascii_btn = gr.Button("Data measures")
    # non_ascii_hist = gr.Plot()
    #
    # gr_ascii_btn.click(non_ascii_check, inputs=[texts_df, text_column], outputs=[non_ascii_hist])

    gr.Markdown("## Explore toxicity")
    # checkbox = gr.Checkbox(value=False, label="Run on full first parquet data (better not)")
    gr_toxicity_btn = gr.Button("Run perpspective API to check toxicity of random samples.")
    toxicity_progress_bar = gr.Label(show_label=False)
    toxicity_hist = gr.Plot()
    with gr.Accordion("Explore examples with toxicity scores:", open=False):
        toxicity_df = gr.DataFrame()
    gr_toxicity_btn.click(
        call_perspective_api,
        inputs=[texts_df, text_column_dropdown, dataset_name, subset_dropdown, split_dropdown],#, checkbox],
        outputs=[toxicity_progress_bar, toxicity_hist, toxicity_df]
    )


demo.launch()