app.py

import gradio as gr
from datasets import load_dataset
import numpy as np
from model2vec import StaticModel
from reach import Reach
from difflib import ndiff
import asyncio

# Load the model at startup
model = StaticModel.from_pretrained("minishlab/M2V_base_output")

# Update default dataset to 'sst2' and set default threshold to 0.9
default_dataset1_name = "sst2"
default_dataset1_split = "train"
default_dataset2_name = "sst2"
default_dataset2_split = "validation"
default_text_column = "sentence"
default_threshold = 0.9

# Load the default datasets at startup
ds_default1 = load_dataset(default_dataset1_name, split=default_dataset1_split)
ds_default2 = load_dataset(default_dataset2_name, split=default_dataset2_split)

def batch_iterable(iterable, batch_size):
    """Helper function to create batches from an iterable."""
    for i in range(0, len(iterable), batch_size):
        yield iterable[i:i + batch_size]

def display_word_differences(x: str, y: str) -> str:
    diff = ndiff(x.split(), y.split())
    return " ".join([word for word in diff if word.startswith(('+', '-'))])

async def compute_embeddings_async(texts, batch_size, progress, desc):
    embeddings = []
    total_batches = (len(texts) + batch_size - 1) // batch_size
    for i, batch_texts in enumerate(batch_iterable(texts, batch_size)):
        batch_embeddings = await asyncio.to_thread(model.encode, batch_texts, show_progressbar=False)
        embeddings.append(batch_embeddings)
        progress((i + 1) / total_batches, desc=desc)
        await asyncio.sleep(0)
    embedding_matrix = np.concatenate(embeddings, axis=0)
    return embedding_matrix

async def deduplicate_async(embedding_matrix: np.ndarray, threshold: float, batch_size: int = 1024, progress=None) -> tuple[np.ndarray, dict[int, int]]:
    """
    Deduplicate embeddings asynchronously.
    """
    progress(0, desc="Building search index...")
    reach = Reach(vectors=embedding_matrix, items=[str(i) for i in range(len(embedding_matrix))])

    deduplicated_indices = set(range(len(embedding_matrix)))
    duplicate_to_original_mapping = {}

    progress(0, desc="Finding nearest neighbors...")
    results = await asyncio.to_thread(reach.nearest_neighbor_threshold,
                                      embedding_matrix,
                                      threshold=threshold,
                                      batch_size=batch_size,
                                      show_progressbar=False)

    total_items = len(embedding_matrix)
    for i, similar_items in enumerate(results):
        if i not in deduplicated_indices:
            continue

        similar_indices = [int(item[0]) for item in similar_items if int(item[0]) != i]

        for sim_idx in similar_indices:
            if sim_idx in deduplicated_indices:
                deduplicated_indices.remove(sim_idx)
                duplicate_to_original_mapping[sim_idx] = i

        if i % 100 == 0:
            progress(i / total_items, desc="Processing duplicates")
            await asyncio.sleep(0)

    progress(1, desc="Processing duplicates")
    return np.array(list(deduplicated_indices)), duplicate_to_original_mapping

async def perform_deduplication(
    deduplication_type,
    dataset1_name,
    dataset1_split,
    dataset1_text_column,
    dataset2_name="",
    dataset2_split="",
    dataset2_text_column="",
    threshold=default_threshold,
    progress=gr.Progress(track_tqdm=True)
):
    try:
        # Convert threshold to float
        threshold = float(threshold)

        # Initialize status message
        status = ""

        if deduplication_type == "Single dataset":
            # Load Dataset 1
            status = "Loading Dataset 1..."
            yield status, ""
            if dataset1_name == default_dataset1_name and dataset1_split == default_dataset1_split:
                ds = ds_default1
            else:
                ds = load_dataset(dataset1_name, split=dataset1_split)

            # Extract texts
            status = "Extracting texts from Dataset 1..."
            yield status, ""
            texts = [example[dataset1_text_column] for example in ds]

            # Compute embeddings
            status = "Computing embeddings for Dataset 1..."
            yield status, ""
            embedding_matrix = await compute_embeddings_async(texts, batch_size=64, progress=progress, desc="Computing embeddings for Dataset 1")

            # Deduplicate
            status = "Deduplicating embeddings..."
            yield status, ""
            deduplicated_indices, duplicate_to_original_mapping = await deduplicate_async(
                embedding_matrix, threshold, progress=progress
            )

            # Prepare the results
            num_duplicates = len(duplicate_to_original_mapping)
            num_total = len(texts)
            num_deduplicated = len(deduplicated_indices)

            result_text = f"**Total documents:** {num_total}\n"
            result_text += f"**Number of duplicates found:** {num_duplicates}\n"
            result_text += f"**Number of unique documents after deduplication:** {num_deduplicated}\n\n"

            # Show deduplicated examples
            if num_duplicates > 0:
                result_text += "**Examples of duplicates found:**\n\n"
                num_examples = min(5, num_duplicates)
                for duplicate_idx, original_idx in list(duplicate_to_original_mapping.items())[:num_examples]:
                    original_text = texts[original_idx]
                    duplicate_text = texts[duplicate_idx]
                    differences = display_word_differences(original_text, duplicate_text)
                    result_text += f"**Original text:**\n{original_text}\n\n"
                    result_text += f"**Duplicate text:**\n{duplicate_text}\n\n"
                    result_text += f"**Differences:**\n{differences}\n"
                    result_text += "-" * 50 + "\n\n"
            else:
                result_text += "No duplicates found."

            # Final status
            status = "Deduplication completed."
            yield status, result_text

        elif deduplication_type == "Cross-dataset":
            # Similar code for cross-dataset deduplication, using async functions
            # Load Dataset 1
            status = "Loading Dataset 1..."
            yield status, ""
            if dataset1_name == default_dataset1_name and dataset1_split == default_dataset1_split:
                ds1 = ds_default1
            else:
                ds1 = load_dataset(dataset1_name, split=dataset1_split)

            # Load Dataset 2
            status = "Loading Dataset 2..."
            yield status, ""
            if dataset2_name == default_dataset2_name and dataset2_split == default_dataset2_split:
                ds2 = ds_default2
            else:
                ds2 = load_dataset(dataset2_name, split=dataset2_split)

            # Extract texts from Dataset 1
            status = "Extracting texts from Dataset 1..."
            yield status, ""
            texts1 = [example[dataset1_text_column] for example in ds1]

            # Extract texts from Dataset 2
            status = "Extracting texts from Dataset 2..."
            yield status, ""
            texts2 = [example[dataset2_text_column] for example in ds2]

            # Compute embeddings for Dataset 1
            status = "Computing embeddings for Dataset 1..."
            yield status, ""
            embedding_matrix1 = await compute_embeddings_async(texts1, batch_size=64, progress=progress, desc="Computing embeddings for Dataset 1")

            # Compute embeddings for Dataset 2
            status = "Computing embeddings for Dataset 2..."
            yield status, ""
            embedding_matrix2 = await compute_embeddings_async(texts2, batch_size=64, progress=progress, desc="Computing embeddings for Dataset 2")

            # Deduplicate across datasets
            status = "Deduplicating embeddings across datasets..."
            yield status, ""
            duplicate_indices_in_ds2, duplicate_to_original_mapping = await deduplicate_across_datasets_async(
                embedding_matrix1, embedding_matrix2, threshold, progress=progress
            )

            num_duplicates = len(duplicate_indices_in_ds2)
            num_total_ds2 = len(texts2)
            num_unique_ds2 = num_total_ds2 - num_duplicates

            result_text = f"**Total documents in {dataset2_name}/{dataset2_split}:** {num_total_ds2}\n"
            result_text += f"**Number of duplicates found in {dataset2_name}/{dataset2_split}:** {num_duplicates}\n"
            result_text += f"**Number of unique documents in {dataset2_name}/{dataset2_split} after deduplication:** {num_unique_ds2}\n\n"

            # Show deduplicated examples
            if num_duplicates > 0:
                result_text += "**Examples of duplicates found in Dataset 2:**\n\n"
                num_examples = min(5, num_duplicates)
                for duplicate_idx in duplicate_indices_in_ds2[:num_examples]:
                    original_idx = duplicate_to_original_mapping[duplicate_idx]
                    original_text = texts1[original_idx]
                    duplicate_text = texts2[duplicate_idx]
                    differences = display_word_differences(original_text, duplicate_text)
                    result_text += f"**Original text (Dataset 1):**\n{original_text}\n\n"
                    result_text += f"**Duplicate text (Dataset 2):**\n{duplicate_text}\n\n"
                    result_text += f"**Differences:**\n{differences}\n"
                    result_text += "-" * 50 + "\n\n"
            else:
                result_text += "No duplicates found."

            # Final status
            status = "Deduplication completed."
            yield status, result_text

    except Exception as e:
        yield f"An error occurred: {e}", ""
        raise e

async def deduplicate_across_datasets_async(embedding_matrix_1: np.ndarray, embedding_matrix_2: np.ndarray, threshold: float, batch_size: int = 1024, progress=None) -> tuple[list[int], dict[int, int]]:
    """
    Deduplicate embeddings across two datasets asynchronously.
    """
    progress(0, desc="Building search index from Dataset 1...")
    reach = Reach(vectors=embedding_matrix_1, items=[str(i) for i in range(len(embedding_matrix_1))])

    duplicate_indices_in_test = []
    duplicate_to_original_mapping = {}

    progress(0, desc="Finding nearest neighbors between datasets...")
    results = await asyncio.to_thread(reach.nearest_neighbor_threshold,
                                      embedding_matrix_2,
                                      threshold=threshold,
                                      batch_size=batch_size,
                                      show_progressbar=False)

    total_items = len(embedding_matrix_2)
    for i, similar_items in enumerate(results):
        similar_indices = [int(item[0]) for item in similar_items if item[1] >= threshold]

        if similar_indices:
            duplicate_indices_in_test.append(i)
            duplicate_to_original_mapping[i] = similar_indices[0]

        if i % 100 == 0:
            progress(i / total_items, desc="Processing duplicates across datasets")