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 # Load the model model = StaticModel.from_pretrained("minishlab/M2V_base_output") # Default parameters default_dataset_name = "sst2" default_dataset1_split = "train" # Default for the first dataset is "train" default_dataset2_split = "test" # Default for the second dataset is "test" default_text_column = "sentence" default_threshold = 0.9 def deduplicate_embeddings( embeddings_a: np.ndarray, embeddings_b: np.ndarray = None, threshold: float = 0.9, batch_size: int = 1024, progress=None ) -> tuple[np.ndarray, dict[int, int]]: """ Deduplicate embeddings within one dataset or across two datasets. :param embeddings_a: Embeddings of Dataset 1. :param embeddings_b: Optional, embeddings of Dataset 2. :param threshold: Similarity threshold for deduplication. :param batch_size: Batch size for similarity computation. :param progress: Gradio progress tracker for feedback. :return: Deduplicated indices and a mapping of removed indices to their original counterparts. """ if embeddings_b is None: reach = Reach(vectors=embeddings_a, items=[str(i) for i in range(len(embeddings_a))]) duplicate_to_original = {} results = reach.nearest_neighbor_threshold( embeddings_a, threshold=threshold, batch_size=batch_size, show_progressbar=False ) for i, similar_items in enumerate(progress.tqdm(results, desc="Processing duplicates", total=len(embeddings_a))): for sim_idx, _ in similar_items: sim_idx = int(sim_idx) if sim_idx != i and sim_idx not in duplicate_to_original: duplicate_to_original[sim_idx] = i deduplicated_indices = set(range(len(embeddings_a))) - set(duplicate_to_original.keys()) return deduplicated_indices, duplicate_to_original else: reach = Reach(vectors=embeddings_a, items=[str(i) for i in range(len(embeddings_a))]) duplicate_indices_in_b = [] duplicate_to_original = {} results = reach.nearest_neighbor_threshold( embeddings_b, threshold=threshold, batch_size=batch_size, show_progressbar=False ) for i, similar_items in enumerate(progress.tqdm(results, desc="Processing duplicates", total=len(embeddings_b))): if similar_items: duplicate_indices_in_b.append(i) duplicate_to_original[i] = int(similar_items[0][0]) return duplicate_indices_in_b, duplicate_to_original def display_word_differences(x: str, y: str) -> str: """ Display the word-level differences between two texts, formatted to avoid misinterpretation of Markdown syntax. :param x: First text. :param y: Second text. :return: A string showing word-level differences, wrapped in a code block. """ diff = ndiff(x.split(), y.split()) formatted_diff = "\n".join(word for word in diff if word.startswith(("+", "-"))) return f"```\n{formatted_diff}\n```" def load_dataset_texts(dataset_name: str, dataset_split: str, text_column: str) -> list[str]: """ Load texts from a specified dataset and split. :param dataset_name: Name of the dataset. :param dataset_split: Split of the dataset (e.g., 'train', 'validation', 'test'). :param text_column: Name of the text column. :return: A list of texts from the dataset. """ ds = load_dataset(dataset_name, split=dataset_split) return [example[text_column] for example in ds] def perform_deduplication( deduplication_type: str, dataset1_name: str, dataset1_split: str, dataset1_text_column: str, dataset2_name: str = "", dataset2_split: str = "", dataset2_text_column: str = "", threshold: float = default_threshold, progress: gr.Progress = gr.Progress(track_tqdm=True) ): """ Perform deduplication on one or two datasets based on the deduplication type. :param deduplication_type: 'Single dataset' or 'Cross-dataset'. :param dataset1_name: Name of the first dataset. :param dataset1_split: Split of the first dataset. :param dataset1_text_column: Text column of the first dataset. :param dataset2_name: Optional, name of the second dataset (for cross-dataset deduplication). :param dataset2_split: Optional, split of the second dataset. :param dataset2_text_column: Optional, text column of the second dataset. :param threshold: Similarity threshold for deduplication. :param progress: Gradio progress tracker. :return: Status updates and result text for the Gradio interface. """ try: threshold = float(threshold) # Load and process Dataset 1 yield "Loading Dataset 1...", "" texts1 = load_dataset_texts(dataset1_name, dataset1_split, dataset1_text_column) yield "Computing embeddings for Dataset 1...", "" embeddings1 = model.encode(texts1, show_progressbar=True) if deduplication_type == "Single dataset": # Deduplicate within Dataset 1 yield "Deduplicating within Dataset 1...", "" deduplicated_indices, duplicate_mapping = deduplicate_embeddings( embeddings1, threshold=threshold, progress=progress ) num_duplicates = len(duplicate_mapping) result_text = ( f"**Total documents:** {len(texts1)}\n\n" f"**Duplicates found:** {num_duplicates}\n\n" f"**Unique documents after deduplication:** {len(deduplicated_indices)}\n\n" + "-" * 50 + "\n\n" ) if num_duplicates > 0: result_text += "**Sample duplicates:**\n\n" for dup_idx, orig_idx in list(duplicate_mapping.items())[:5]: orig_text = texts1[orig_idx] dup_text = texts1[dup_idx] differences = display_word_differences(orig_text, dup_text) result_text += ( f"**Original:**\n{orig_text}\n\n" f"**Duplicate:**\n{dup_text}\n\n" f"**Differences:**\n{differences}\n" + "-" * 50 + "\n\n" ) else: result_text += "No duplicates found." yield "Deduplication completed.", result_text else: # Load and process Dataset 2 yield "Loading Dataset 2...", "" texts2 = load_dataset_texts(dataset2_name, dataset2_split, dataset2_text_column) yield "Computing embeddings for Dataset 2...", "" embeddings2 = model.encode(texts2, show_progressbar=True) # Deduplicate Dataset 2 against Dataset 1 yield "Deduplicating Dataset 2 against Dataset 1...", "" duplicate_indices, duplicate_mapping = deduplicate_embeddings( embeddings1, embeddings_b=embeddings2, threshold=threshold, progress=progress ) num_duplicates = len(duplicate_indices) result_text = ( f"**Total documents in {dataset2_name}/{dataset2_split}:** {len(texts2)}\n\n" f"**Duplicates found in Dataset 2:** {num_duplicates}\n\n" f"**Unique documents after deduplication:** {len(texts2) - num_duplicates}\n\n" + "-" * 50 + "\n\n" ) if num_duplicates > 0: result_text += "**Sample duplicates from Dataset 2:**\n\n" for idx in duplicate_indices[:5]: orig_text = texts1[duplicate_mapping[idx]] dup_text = texts2[idx] differences = display_word_differences(orig_text, dup_text) result_text += ( f"**Original (Dataset 1):**\n{orig_text}\n\n" f"**Duplicate (Dataset 2):**\n{dup_text}\n\n" f"**Differences:**\n{differences}\n" + "-" * 50 + "\n\n" ) else: result_text += "No duplicates found." yield "Deduplication completed.", result_text except Exception as e: yield f"An error occurred: {e}", "" raise e # Gradio app with stop button support with gr.Blocks(theme=gr.themes.Ocean(), css="#status_output { height: 50px; overflow: auto; }") as demo: gr.Markdown("# Semantic Deduplication") gr.Markdown(""" This demo showcases semantic deduplication using Model2Vec for HuggingFace datasets. It can be used to identify duplicate texts within a single dataset or across two datasets. You can adjust the similarity threshold to control the strictness of the deduplication.\n NOTE: this demo runs on a free CPU backend, so it may be slow for large datasets. For faster results, please run the code locally. """) deduplication_type = gr.Radio( choices=["Cross-dataset", "Single dataset"], # Swapped "Cross-dataset" to the left label="Deduplication Type", value="Cross-dataset", # Set "Cross-dataset" as the default value ) with gr.Row(): dataset1_name = gr.Textbox(value=default_dataset_name, label="Dataset 1 Name") dataset1_split = gr.Textbox(value=default_dataset1_split, label="Dataset 1 Split") # Default split is "train" dataset1_text_column = gr.Textbox(value=default_text_column, label="Text Column Name") dataset2_inputs = gr.Column(visible=True) # Make dataset2_inputs visible by default with dataset2_inputs: with gr.Row(): dataset2_name = gr.Textbox(value=default_dataset_name, label="Dataset 2 Name") dataset2_split = gr.Textbox(value=default_dataset2_split, label="Dataset 2 Split") # Default split is "test" dataset2_text_column = gr.Textbox(value=default_text_column, label="Text Column Name") threshold = gr.Slider(0.0, 1.0, value=default_threshold, label="Similarity Threshold") with gr.Row(): # Placing the button in the same row for better alignment compute_button = gr.Button("Deduplicate") status_output = gr.Markdown(elem_id="status_output") result_output = gr.Markdown() def update_visibility(choice: str): return gr.update(visible=choice == "Cross-dataset") deduplication_type.change(update_visibility, inputs=deduplication_type, outputs=dataset2_inputs) compute_button.click( fn=perform_deduplication, inputs=[ deduplication_type, dataset1_name, dataset1_split, dataset1_text_column, dataset2_name, dataset2_split, dataset2_text_column, threshold, ], outputs=[status_output, result_output], ) demo.launch()