File size: 13,448 Bytes
5ae3e12
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0f8853a
5ae3e12
 
 
 
 
 
 
0f8853a
 
 
 
 
 
 
 
 
5ae3e12
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0f8853a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5ae3e12
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0f8853a
5ae3e12
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0f8853a
 
5ae3e12
 
0f8853a
 
5ae3e12
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import hashlib
import json
import multiprocessing as mp
import re
import sqlite3
from collections import Counter, defaultdict
from pathlib import Path
from typing import Dict, Iterable, Iterator, List, Optional

from tqdm import tqdm


TOKEN_PATTERN = re.compile(r"\w+|[^\w\s]", re.UNICODE)
CODE_PATTERN = re.compile(
    r"(\bdef\b|\bclass\b|\bimport\b|\breturn\b|=>|function\s+\w+|public\s+class|#include|```)",
    re.IGNORECASE,
)
MIN_CODE_SIGNAL_RE = re.compile(r"(\bdef\s+|\bclass\s+|\bimport\s+|=|\breturn\s+|\bfor\s+|\bif\s+)")
EXPLANATION_PATTERN = re.compile(
    r"\b(explain|because|algorithm|steps|approach|complexity|solution)\b", re.IGNORECASE
)
PROBLEM_PROMPT_RE = re.compile(
    r"\b(solve|given|find|compute|return|input|output|problem|algorithm|task|challenge)\b",
    re.IGNORECASE,
)
YAML_FRONTMATTER_RE = re.compile(r"^\s*---\s*\n.*?\n---\s*", re.DOTALL)
FENCE_LINE_RE = re.compile(r"^\s*```(?:[a-zA-Z0-9_+-]+)?\s*$")
CODEBLOCK_DIRECTIVE_RE = re.compile(r"^\s*code-block::\s*\w*\s*$", re.IGNORECASE)
CLI_NOISE_LINE_RE = re.compile(
    r"^\s*(\[[^\]]+\]\s*)?(INFO|WARNING|ERROR|DEBUG|TRACE)\b|"
    r"^\s*(PS\s+[A-Za-z]:\\|[A-Za-z]:\\[^>]*>)|"
    r"^\s*(Traceback \(most recent call last\):|File \".*\", line \d+|Exception:|RuntimeError:|ValueError:|TypeError:)",
    re.IGNORECASE,
)


def estimate_tokens(text: str) -> int:
    if not text:
        return 0
    return len(TOKEN_PATTERN.findall(text))


def normalize_text(text: str) -> str:
    if text is None:
        return ""
    text = str(text).replace("\x00", "")
    text = text.replace("\r\n", "\n").replace("\r", "\n")
    text = "".join(ch for ch in text if ch == "\n" or ch == "\t" or ord(ch) >= 32)
    lines = [line.rstrip() for line in text.split("\n")]
    return "\n".join(lines).strip()


def clean_response_text(response: str) -> str:
    text = normalize_text(response)
    if not text:
        return ""

    # Remove YAML front matter and markdown wrappers.
    text = YAML_FRONTMATTER_RE.sub("", text)
    kept_lines: List[str] = []
    for line in text.split("\n"):
        if FENCE_LINE_RE.match(line):
            continue
        if CODEBLOCK_DIRECTIVE_RE.match(line):
            continue
        if CLI_NOISE_LINE_RE.match(line):
            continue
        kept_lines.append(line)

    text = "\n".join(kept_lines)
    text = text.replace("```python", "").replace("```py", "").replace("```", "")

    # Normalize indentation and drop leading/trailing blank lines.
    normalized_lines = [ln.rstrip() for ln in text.split("\n")]
    while normalized_lines and not normalized_lines[0].strip():
        normalized_lines.pop(0)
    while normalized_lines and not normalized_lines[-1].strip():
        normalized_lines.pop()
    if not normalized_lines:
        return ""

    # Keep indentation consistent: convert tabs to 4 spaces.
    normalized_lines = [ln.replace("\t", "    ") for ln in normalized_lines]
    text = "\n".join(normalized_lines)
    return text


def _ascii_ratio(text: str) -> float:
    if not text:
        return 1.0
    ascii_count = sum(1 for c in text if ord(c) < 128)
    return ascii_count / len(text)


def _response_is_valid(response: str) -> bool:
    if not response:
        return False
    if CODE_PATTERN.search(response):
        return True
    if EXPLANATION_PATTERN.search(response):
        return True
    return False


def _response_has_code(response: str) -> bool:
    return bool(
        re.search(
            r"(\bdef\b|\bclass\b|\breturn\b|\bimport\b|```|function\s+\w+|public\s+class|#include|SELECT\s+)",
            response,
            re.IGNORECASE,
        )
    )


def clean_record(
    record: Dict[str, str],
    *,
    min_tokens: int = 10,
    max_tokens: int = 2048,
) -> Optional[Dict[str, str]]:
    instruction = normalize_text(record.get("instruction", ""))
    response = clean_response_text(record.get("response", ""))
    source = normalize_text(record.get("_source", "unknown"))
    category = normalize_text(record.get("_category", ""))
    if not category:
        src_low = source.lower()
        if any(k in src_low for k in ("codealpaca", "evol", "ultrachat", "openhermes", "orca")):
            category = "instruction"
        elif any(
            k in src_low
            for k in (
                "leetcode",
                "contest",
                "mbpp",
                "humaneval",
                "apps",
                "codeforces",
                "problem",
                "codesearchnet_problem",
            )
        ):
            category = "problem"
        else:
            category = "structured"

    if not instruction or not response:
        return None
    if len(response) < 40:
        return None
    if _ascii_ratio(instruction + response) < 0.85:
        return None
    if not MIN_CODE_SIGNAL_RE.search(response):
        return None
    if not _response_is_valid(response):
        return None
    if category == "problem":
        if len(instruction) <= 50:
            return None
        if not PROBLEM_PROMPT_RE.search(instruction):
            return None
        if not _response_has_code(response):
            return None
        # Problem solutions must include code, not explanation-only text.
        if EXPLANATION_PATTERN.search(response) and not CODE_PATTERN.search(response):
            return None

    total_tokens = estimate_tokens(instruction) + estimate_tokens(response)
    if total_tokens < min_tokens or total_tokens > max_tokens:
        return None

    return {
        "instruction": instruction,
        "response": response,
        "_source": source,
        "_category": category,
        "_tokens": total_tokens,
    }


def _iter_jsonl(path: Path) -> Iterable[Dict[str, str]]:
    with path.open("r", encoding="utf-8") as f:
        for line in f:
            line = line.strip()
            if not line:
                continue
            try:
                yield json.loads(line)
            except json.JSONDecodeError:
                continue


def _clean_record_worker(payload: Dict[str, object]) -> Optional[Dict[str, str]]:
    record = payload["record"]
    min_tokens = int(payload["min_tokens"])
    max_tokens = int(payload["max_tokens"])
    return clean_record(record, min_tokens=min_tokens, max_tokens=max_tokens)


def iter_cleaned_records(
    path: Path,
    *,
    min_tokens: int,
    max_tokens: int,
    num_workers: int = 1,
    batch_size: int = 2000,
) -> Iterator[Dict[str, str]]:
    if num_workers <= 1:
        for record in _iter_jsonl(path):
            cleaned = clean_record(record, min_tokens=min_tokens, max_tokens=max_tokens)
            if cleaned is not None:
                yield cleaned
        return

    pool = mp.Pool(processes=num_workers)
    try:
        batch: List[Dict[str, str]] = []
        for record in _iter_jsonl(path):
            batch.append(record)
            if len(batch) < batch_size:
                continue
            payloads = [
                {"record": r, "min_tokens": min_tokens, "max_tokens": max_tokens} for r in batch
            ]
            for cleaned in pool.imap_unordered(_clean_record_worker, payloads, chunksize=64):
                if cleaned is not None:
                    yield cleaned
            batch.clear()

        if batch:
            payloads = [{"record": r, "min_tokens": min_tokens, "max_tokens": max_tokens} for r in batch]
            for cleaned in pool.imap_unordered(_clean_record_worker, payloads, chunksize=64):
                if cleaned is not None:
                    yield cleaned
    finally:
        pool.close()
        pool.join()


def _remove_sqlite_artifacts(sqlite_path: Path) -> None:
    if sqlite_path.exists():
        sqlite_path.unlink()
    for suffix in ("-wal", "-shm"):
        p = sqlite_path.with_name(sqlite_path.name + suffix)
        if p.exists():
            p.unlink()


def _open_dedupe_db(sqlite_path: Path):
    sqlite_path = sqlite_path.resolve()
    sqlite_path.parent.mkdir(parents=True, exist_ok=True)
    _remove_sqlite_artifacts(sqlite_path)
    conn = sqlite3.connect(str(sqlite_path))
    conn.execute("PRAGMA journal_mode=WAL;")
    conn.execute("CREATE TABLE IF NOT EXISTS seen_hashes (h TEXT PRIMARY KEY)")
    return conn


def _is_duplicate(conn, instruction: str, response: str) -> bool:
    digest = hashlib.sha256(f"{instruction}||{response}".encode("utf-8")).hexdigest()
    try:
        conn.execute("INSERT INTO seen_hashes(h) VALUES (?)", (digest,))
        return False
    except sqlite3.IntegrityError:
        return True


def build_balanced_dataset(
    input_paths: List[Path],
    output_path: Path,
    *,
    target_size: int = 1_000_000,
    min_tokens: int = 10,
    max_tokens: int = 2048,
    category_weights: Optional[Dict[str, float]] = None,
    sqlite_path: Optional[Path] = None,
    num_workers: int = 1,
) -> Dict[str, object]:
    output_path.parent.mkdir(parents=True, exist_ok=True)
    if sqlite_path is None:
        sqlite_path = output_path.parent / "dedupe_hashes.sqlite"
    conn = _open_dedupe_db(sqlite_path)

    weights = category_weights or {"instruction": 0.60, "structured": 0.30, "problem": 0.10}
    target_by_cat = {k: int(target_size * v) for k, v in weights.items()}
    target_by_cat["problem"] = target_size - target_by_cat["instruction"] - target_by_cat["structured"]

    grouped_paths: Dict[str, List[Path]] = defaultdict(list)
    for path in input_paths:
        if not path.exists():
            continue
        name = path.stem
        if "codealpaca" in name or "evol" in name or "ultrachat" in name or "openhermes" in name:
            grouped_paths["instruction"].append(path)
        elif any(
            k in name
            for k in (
                "leetcode",
                "contest",
                "problem",
                "mbpp",
                "humaneval",
                "apps",
                "codeforces",
            )
        ):
            grouped_paths["problem"].append(path)
        else:
            grouped_paths["structured"].append(path)

    source_counter = Counter()
    category_counter = Counter()
    total_tokens = 0
    total_kept = 0

    def try_write(cleaned: Dict[str, str], out_f, enforce_category_target: bool) -> bool:
        nonlocal total_kept, total_tokens
        category = cleaned["_category"]
        if enforce_category_target and category_counter[category] >= target_by_cat.get(category, 0):
            return False
        if _is_duplicate(conn, cleaned["instruction"], cleaned["response"]):
            return False
        source = cleaned["_source"]
        tokens = int(cleaned["_tokens"])
        category_counter[category] += 1
        source_counter[source] += 1
        total_tokens += tokens
        total_kept += 1
        out_f.write(
            json.dumps(
                {"instruction": cleaned["instruction"], "response": cleaned["response"]},
                ensure_ascii=False,
            )
            + "\n"
        )
        return True

    with output_path.open("w", encoding="utf-8") as out_f:
        # Phase 1: enforce 60/30/10 quotas.
        for category in ("instruction", "structured", "problem"):
            if category not in grouped_paths:
                continue
            for path in grouped_paths[category]:
                cleaned_iter = iter_cleaned_records(
                    path,
                    min_tokens=min_tokens,
                    max_tokens=max_tokens,
                    num_workers=num_workers,
                )
                for cleaned in tqdm(cleaned_iter, desc=f"balance1:{path.name}", unit="rows"):
                    if total_kept >= target_size or category_counter[category] >= target_by_cat[category]:
                        break
                    try_write(cleaned, out_f, enforce_category_target=True)
                conn.commit()
                if total_kept >= target_size or category_counter[category] >= target_by_cat[category]:
                    continue

        # Phase 2: fill remaining slots from all categories while preserving dedupe.
        if total_kept < target_size:
            for path in input_paths:
                if not path.exists():
                    continue
                cleaned_iter = iter_cleaned_records(
                    path,
                    min_tokens=min_tokens,
                    max_tokens=max_tokens,
                    num_workers=num_workers,
                )
                for cleaned in tqdm(cleaned_iter, desc=f"balance2:{path.name}", unit="rows"):
                    if total_kept >= target_size:
                        break
                    try_write(cleaned, out_f, enforce_category_target=False)
                conn.commit()
                if total_kept >= target_size:
                    break

    conn.close()
    avg_len = round((total_tokens / total_kept), 2) if total_kept else 0.0
    raw_converted = category_counter["structured"] + category_counter["problem"]
    ratio = {
        "instruction_pct": round(100.0 * category_counter["instruction"] / max(total_kept, 1), 2),
        "raw_converted_pct": round(100.0 * raw_converted / max(total_kept, 1), 2),
    }

    return {
        "total_samples": total_kept,
        "avg_length_tokens": avg_len,
        "source_breakdown": dict(source_counter),
        "category_breakdown": dict(category_counter),
        "instruction_vs_raw_ratio": ratio,
        "targets": target_by_cat,
    }