File size: 19,835 Bytes
25d2eb7
2827b8a
 
39a5b1c
2827b8a
7a1cd7a
a81fb12
95530b9
39a5b1c
f5eb405
95530b9
 
 
3b4c438
 
f5eb405
95530b9
 
 
 
c58907b
 
24f7d5b
ed5b7bd
 
 
 
 
 
 
 
 
 
95530b9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
58d8f1a
7a1cd7a
ed5b7bd
 
 
 
 
 
 
 
7a1cd7a
73a84b9
 
 
24f7d5b
ed5b7bd
 
 
 
 
 
 
 
95530b9
 
7a1cd7a
4f0286f
24f7d5b
 
 
 
 
 
 
 
 
2827b8a
ed5b7bd
 
 
 
 
 
 
 
 
 
 
 
 
 
f5eb405
 
3bd0812
95530b9
 
 
 
f39d105
24f7d5b
f5eb405
95530b9
 
 
 
c58907b
 
95530b9
 
2a0be82
 
95530b9
c58907b
3bd0812
5422464
95530b9
 
 
 
 
 
 
 
 
 
 
5422464
 
3bd0812
95530b9
c58907b
95530b9
 
 
 
 
f39d105
24f7d5b
95530b9
 
 
 
c58907b
 
95530b9
 
2a0be82
 
95530b9
c58907b
39a5b1c
 
95530b9
 
 
 
 
 
 
 
 
 
 
39a5b1c
 
 
95530b9
f5eb405
6b0e834
39a5b1c
c58907b
 
72c7e2c
1282e7b
e49e0e9
24f7d5b
d54c792
 
 
 
24f7d5b
4f0286f
 
 
 
c58907b
4f0286f
 
 
95530b9
 
4f0286f
 
 
 
 
 
95530b9
 
4f0286f
 
95530b9
d54c792
1a5f99b
c58907b
4f0286f
24f7d5b
95530b9
4f0286f
95530b9
4f0286f
 
 
 
 
 
 
 
 
 
 
c58907b
4f0286f
c58907b
4f0286f
 
72c7e2c
4f0286f
24f7d5b
ed5b7bd
 
72c7e2c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ed5b7bd
72c7e2c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ed5b7bd
72c7e2c
 
 
 
 
ed5b7bd
72c7e2c
 
 
 
 
 
 
 
 
 
 
 
 
 
ed5b7bd
72c7e2c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ed5b7bd
72c7e2c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ed5b7bd
72c7e2c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ed5b7bd
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
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_dataset_split = "train"
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').
    :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"
            )

            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"
            )

            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(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=["Single dataset", "Cross-dataset"],
        label="Deduplication Type",
        value="Single dataset",
    )

    with gr.Row():
        dataset1_name = gr.Textbox(value=default_dataset_name, label="Dataset 1 Name")
        dataset1_split = gr.Textbox(value=default_dataset_split, label="Dataset 1 Split")
        dataset1_text_column = gr.Textbox(value=default_text_column, label="Text Column Name")

    dataset2_inputs = gr.Column(visible=False)
    with dataset2_inputs:
        gr.Markdown("### Dataset 2")
        with gr.Row():
            dataset2_name = gr.Textbox(value=default_dataset_name, label="Dataset 2 Name")
            dataset2_split = gr.Textbox(value=default_dataset_split, label="Dataset 2 Split")
            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")
    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()



# 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_dataset_split = "train"
# 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."""
#     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 word-level differences between two texts, avoiding Markdown issues."""
#     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."""
#     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."""
#     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"
#             )

#             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"
#             )

#             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(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=["Single dataset", "Cross-dataset"],
#         label="Deduplication Type",
#         value="Single dataset",
#     )

#     with gr.Row():
#         dataset1_name = gr.Textbox(value=default_dataset_name, label="Dataset 1 Name")
#         dataset1_split = gr.Textbox(value=default_dataset_split, label="Dataset 1 Split")
#         dataset1_text_column = gr.Textbox(value=default_text_column, label="Text Column Name")

#     dataset2_inputs = gr.Column(visible=False)
#     with dataset2_inputs:
#         gr.Markdown("### Dataset 2")
#         with gr.Row():
#             dataset2_name = gr.Textbox(value=default_dataset_name, label="Dataset 2 Name")
#             dataset2_split = gr.Textbox(value=default_dataset_split, label="Dataset 2 Split")
#             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")
#     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()