File size: 16,669 Bytes
4e8a7b5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Dokime — HuggingFace Spaces demo.

A Gradio app showcasing Dokime's core data curation features:
filtering, quality scoring, and semantic search.
"""

from __future__ import annotations

import json
import time
from typing import Any

import gradio as gr

# ---------------------------------------------------------------------------
# Dokime imports
# ---------------------------------------------------------------------------
from dokime.core.filters import (
    AlphaFilter,
    LengthFilter,
    RepetitionFilter,
    WhitespaceFilter,
)
from dokime.core.pipeline import Pipeline
from dokime.quality.dedup import ExactDedup
from dokime.quality.scoring import QualityScorer

# ---------------------------------------------------------------------------
# Sample dataset — pre-loaded so users can click "Try it" immediately
# ---------------------------------------------------------------------------
SAMPLE_JSONL = r"""{"text": "Machine learning is a subfield of artificial intelligence that gives computers the ability to learn without being explicitly programmed. It focuses on the development of algorithms that can access data and use it to learn for themselves. The process begins with observations or data, such as examples, direct experience, or instruction, in order to look for patterns in data and make better decisions in the future."}
{"text": "The   quick    brown    fox   jumps    over   the   lazy    dog.   Too   much   whitespace   here.   Way   too   much.    Formatting   is   broken.    Lots    of    extra    spaces    everywhere    in    this    document.    Not    great    quality."}
{"text": "buy cheap viagra online best price discount pharmacy click here now limited time offer buy buy buy click click click spam spam spam buy cheap viagra online best price discount pharmacy click here now limited time offer buy buy buy click click click"}
{"text": "Deep learning architectures, particularly transformer models, have revolutionized natural language processing. Models like BERT, GPT, and T5 demonstrate that pre-training on large corpora followed by task-specific fine-tuning yields state-of-the-art results across a wide range of benchmarks. These advances have enabled practical applications from machine translation to code generation."}
{"text": "ab"}
{"text": "Data curation is the process of organizing, integrating, and maintaining data collections to ensure they meet the needs of downstream consumers. For machine learning, good data curation involves removing duplicates, filtering low-quality samples, balancing class distributions, and ensuring data provenance. Dokime provides tools for all of these tasks in a single Python package."}
{"text": "The quick brown fox jumps over the lazy dog near the river bank on a sunny afternoon while the birds sing melodiously in the tall oak trees swaying gently in the warm summer breeze that carries the scent of wildflowers across the rolling green meadows stretching toward the distant purple mountains."}
{"text": "12345 67890 !@#$% ^&*() 98765 43210 !@#$% ^&*() 12345 67890 !@#$% ^&*() numbers and symbols only 54321 09876 !@#$% ^&*()"}
{"text": "Data curation is the process of organizing, integrating, and maintaining data collections to ensure they meet the needs of downstream consumers. For machine learning, good data curation involves removing duplicates, filtering low-quality samples, balancing class distributions, and ensuring data provenance. Dokime provides tools for all of these tasks in a single Python package."}
{"text": "Retrieval-augmented generation (RAG) combines information retrieval with language generation to produce more factual and grounded outputs. By retrieving relevant passages from a knowledge base before generating a response, RAG systems can reduce hallucination and provide up-to-date information that was not seen during pre-training. This approach is becoming standard in enterprise AI deployments."}""".strip()


# ---------------------------------------------------------------------------
# Helper: parse JSONL input (tolerant)
# ---------------------------------------------------------------------------
def _parse_jsonl(raw: str) -> list[dict[str, Any]]:
    """Parse JSONL text into a list of dicts. Skips blank / unparseable lines."""
    docs: list[dict[str, Any]] = []
    for line in raw.strip().splitlines():
        line = line.strip()
        if not line:
            continue
        try:
            obj = json.loads(line)
            if isinstance(obj, dict):
                docs.append(obj)
        except json.JSONDecodeError:
            # Treat bare text lines as {"text": line}
            docs.append({"text": line})
    return docs


def _docs_to_jsonl(docs: list[dict[str, Any]]) -> str:
    return "\n".join(json.dumps(d, ensure_ascii=False) for d in docs)


# ===================================================================
# Tab 1: Curate
# ===================================================================
def run_curate(
    raw_jsonl: str,
    min_length: int,
    max_whitespace: float,
    max_repetition: float,
    enable_dedup: bool,
) -> tuple[str, str]:
    """Build and run a Dokime pipeline on the pasted JSONL data."""
    docs = _parse_jsonl(raw_jsonl)
    if not docs:
        return "", "No valid documents found in input."

    # Build pipeline
    pipeline = Pipeline("spaces-demo")
    pipeline.add_filter(LengthFilter(min_length=min_length))
    pipeline.add_filter(WhitespaceFilter(max_whitespace_ratio=max_whitespace))
    pipeline.add_filter(RepetitionFilter(max_repetition_ratio=max_repetition))
    if enable_dedup:
        pipeline.add_filter(ExactDedup())

    # Run filter chain manually to collect per-filter stats
    filter_stats: dict[str, int] = {f.name(): 0 for f in pipeline.filters}
    kept: list[dict[str, Any]] = []
    start = time.time()

    for sample in docs:
        passed = True
        for f in pipeline.filters:
            if not f.filter(sample):
                filter_stats[f.name()] += 1
                passed = False
                break
        if passed:
            kept.append(sample)

    elapsed = time.time() - start
    total_read = len(docs)
    total_kept = len(kept)
    total_removed = total_read - total_kept
    removal_pct = (total_removed / total_read * 100) if total_read else 0

    # Build stats markdown table
    lines = [
        "| Metric | Value |",
        "|---|---|",
        f"| Documents read | {total_read} |",
        f"| Documents kept | {total_kept} |",
        f"| Documents removed | {total_removed} |",
        f"| Removal rate | {removal_pct:.1f}% |",
        f"| Elapsed | {elapsed:.3f}s |",
        "",
        "**Per-filter removals:**",
        "",
        "| Filter | Removed |",
        "|---|---|",
    ]
    for fname, count in filter_stats.items():
        lines.append(f"| `{fname}` | {count} |")

    return _docs_to_jsonl(kept), "\n".join(lines)


# ===================================================================
# Tab 2: Score
# ===================================================================
def run_score(raw_text: str) -> str:
    """Score each document with Dokime's QualityScorer."""
    docs = _parse_jsonl(raw_text)
    if not docs:
        return "No valid documents found in input."

    scorer = QualityScorer()

    header = (
        "| # | text (first 80 chars) | _char_entropy | _alpha_ratio | "
        "_whitespace_ratio | _special_ratio | _avg_word_length | _quality_score |"
    )
    sep = "|---|---|---|---|---|---|---|---|"
    rows = [header, sep]

    for i, doc in enumerate(docs, 1):
        scored = scorer.score(doc)
        preview = scored.get("text", "")[:80].replace("|", "\\|").replace("\n", " ")
        rows.append(
            f"| {i} | {preview} | {scored['_char_entropy']} | "
            f"{scored['_alpha_ratio']} | {scored['_whitespace_ratio']} | "
            f"{scored['_special_ratio']} | {scored['_avg_word_length']} | "
            f"**{scored['_quality_score']}** |"
        )

    return "\n".join(rows)


# ===================================================================
# Tab 3: Search
# ===================================================================

# Lazy-loaded globals for search (avoid loading model on startup)
_search_model = None
_search_index = None
_search_docs: list[dict[str, Any]] = []


def _ensure_search_model():
    global _search_model
    if _search_model is None:
        from dokime.embeddings.compute import EmbeddingModel

        _search_model = EmbeddingModel("all-MiniLM-L6-v2")
    return _search_model


def run_search(raw_jsonl: str, query: str, top_k: int) -> str:
    """Embed the dataset, build a FAISS index, and search."""
    global _search_index, _search_docs

    docs = _parse_jsonl(raw_jsonl)
    if not docs:
        return "No valid documents found in input."
    if not query.strip():
        return "Please enter a search query."

    model = _ensure_search_model()

    # Recompute index if the docs changed
    texts = [d.get("text", "") for d in docs]
    import numpy as np

    embeddings = model.encode_texts(texts, show_progress=False)

    from dokime.embeddings.search import EmbeddingIndex

    index = EmbeddingIndex(embeddings, docs)
    results = index.search(query, model, k=min(top_k, len(docs)))

    if not results:
        return "No results found."

    header = "| Rank | Score | Text (first 120 chars) |"
    sep = "|---|---|---|"
    rows = [header, sep]
    for rank, r in enumerate(results, 1):
        preview = r.document.get("text", "")[:120].replace("|", "\\|").replace("\n", " ")
        rows.append(f"| {rank} | {r.score:.4f} | {preview} |")

    return "\n".join(rows)


# ===================================================================
# UI — dark theme, polished layout
# ===================================================================

DESCRIPTION = """\
# Dokime

**pytest for your training data.**

Filter junk, deduplicate, score quality, compute embeddings, and search your \
dataset -- all from one `pip install`.
"""

CSS = """
.dokime-header {
    text-align: center;
    margin-bottom: 0.5em;
}
footer { display: none !important; }
"""

with gr.Blocks(
    theme=gr.themes.Base(
        primary_hue=gr.themes.colors.purple,
        secondary_hue=gr.themes.colors.teal,
        neutral_hue=gr.themes.colors.gray,
        font=gr.themes.GoogleFont("Inter"),
    ),
    css=CSS,
    title="Dokime — Data Curation Toolkit",
) as demo:

    gr.Markdown(DESCRIPTION, elem_classes=["dokime-header"])

    # ----- Tab 1: Curate -----
    with gr.Tab("Curate"):
        gr.Markdown(
            "Paste JSONL data below (each line: `{\"text\": \"...\"}`) and "
            "configure filters. Click **Run Pipeline** to see what Dokime keeps "
            "and what it removes."
        )
        with gr.Row():
            with gr.Column(scale=2):
                curate_input = gr.Textbox(
                    label="Input (JSONL)",
                    lines=12,
                    max_lines=30,
                    placeholder='{"text": "Your document here..."}\n{"text": "Another document..."}',
                )
                curate_sample_btn = gr.Button(
                    "Load sample dataset", variant="secondary", size="sm"
                )
            with gr.Column(scale=1):
                min_length = gr.Slider(
                    minimum=0,
                    maximum=500,
                    value=20,
                    step=5,
                    label="Min length (chars)",
                )
                max_whitespace = gr.Slider(
                    minimum=0.1,
                    maximum=1.0,
                    value=0.5,
                    step=0.05,
                    label="Max whitespace ratio",
                )
                max_repetition = gr.Slider(
                    minimum=0.05,
                    maximum=1.0,
                    value=0.3,
                    step=0.05,
                    label="Max repetition ratio",
                )
                enable_dedup = gr.Checkbox(label="Exact deduplication", value=True)
                curate_btn = gr.Button("Run Pipeline", variant="primary")

        curate_stats = gr.Markdown(label="Statistics")
        curate_output = gr.Textbox(
            label="Output (kept documents, JSONL)", lines=10, max_lines=30
        )

        curate_sample_btn.click(
            fn=lambda: SAMPLE_JSONL, outputs=[curate_input]
        )
        curate_btn.click(
            fn=run_curate,
            inputs=[curate_input, min_length, max_whitespace, max_repetition, enable_dedup],
            outputs=[curate_output, curate_stats],
        )

    # ----- Tab 2: Score -----
    with gr.Tab("Score"):
        gr.Markdown(
            "Paste documents below (JSONL or one text per line) and click "
            "**Score Quality** to see per-document quality signals."
        )
        score_input = gr.Textbox(
            label="Input (JSONL or plain text, one doc per line)",
            lines=10,
            max_lines=30,
            placeholder='{"text": "Your document here..."}\n{"text": "Another document..."}',
        )
        with gr.Row():
            score_sample_btn = gr.Button(
                "Load sample dataset", variant="secondary", size="sm"
            )
            score_btn = gr.Button("Score Quality", variant="primary")
        score_output = gr.Markdown(label="Quality Scores")

        score_sample_btn.click(fn=lambda: SAMPLE_JSONL, outputs=[score_input])
        score_btn.click(fn=run_score, inputs=[score_input], outputs=[score_output])

    # ----- Tab 3: Search -----
    with gr.Tab("Search"):
        gr.Markdown(
            "Paste a JSONL dataset, enter a natural-language query, and "
            "Dokime will rank documents by semantic similarity using "
            "sentence-transformers + FAISS.\n\n"
            "> **Note:** The first search loads the `all-MiniLM-L6-v2` model "
            "(~80 MB). This may take 10-30 seconds on a cold start."
        )
        search_input = gr.Textbox(
            label="Dataset (JSONL)",
            lines=10,
            max_lines=30,
            placeholder='{"text": "Your document here..."}\n{"text": "Another document..."}',
        )
        with gr.Row():
            search_query = gr.Textbox(
                label="Search query",
                placeholder="e.g. retrieval augmented generation",
                scale=3,
            )
            search_k = gr.Slider(
                minimum=1, maximum=20, value=5, step=1, label="Top K", scale=1
            )
        with gr.Row():
            search_sample_btn = gr.Button(
                "Load sample dataset", variant="secondary", size="sm"
            )
            search_btn = gr.Button("Search", variant="primary")
        search_output = gr.Markdown(label="Search Results")

        search_sample_btn.click(fn=lambda: SAMPLE_JSONL, outputs=[search_input])
        search_btn.click(
            fn=run_search,
            inputs=[search_input, search_query, search_k],
            outputs=[search_output],
        )

    # ----- Tab 4: About -----
    with gr.Tab("About"):
        gr.Markdown(
            """
## What is Dokime?

Dokime is an open-source Python toolkit for curating ML training data. It
provides heuristic filters, deduplication (exact, fuzzy, and semantic),
quality scoring, embedding computation, semantic search, and outlier
detection -- all from a single `pip install`.

### Core capabilities

- **12 built-in heuristic filters** -- length, whitespace, repetition, special
  characters, alpha ratio, URLs, stopwords, language, regex, and more
- **3 deduplication strategies** -- exact (SHA-256), fuzzy (MinHash-LSH), and
  semantic (embedding cosine similarity)
- **Quality scoring** -- per-document signals including character entropy,
  alpha ratio, composite quality score
- **Embeddings & search** -- sentence-transformer embeddings, FAISS
  nearest-neighbor search, anomaly/outlier detection
- **Pipeline orchestration** -- chain filters in code or YAML, get per-filter
  removal stats
- **CLI + Python SDK** -- seven commands, zero boilerplate

### Links

| | |
|---|---|
| GitHub | [github.com/dokime-ai/dokime](https://github.com/dokime-ai/dokime) |
| PyPI | [pypi.org/project/dokime](https://pypi.org/project/dokime/) |
| Docs | [dokime-ai.github.io/dokime](https://dokime-ai.github.io/dokime) |

### Install

```bash
pip install dokime
```

For the full toolkit (embeddings, fuzzy dedup, language detection):

```bash
pip install "dokime[all]"
```

### License

Apache 2.0
"""
        )

if __name__ == "__main__":
    demo.launch()