| |
| """Streaming integrity and content checks for the expansion. |
| |
| This is intentionally independent of the writers' in-memory uniqueness checks: |
| it reads the files that will actually be uploaded. |
| """ |
| from __future__ import annotations |
|
|
| import csv |
| import hashlib |
| import json |
| import re |
| from collections import Counter |
| from pathlib import Path |
|
|
| from common import REPO, OUT, text_sha256 |
| from generate_creative import RHYME_PAIRS, digits |
|
|
| EXPECTED = { |
| "common_sense_v2.csv": 100_000, |
| "small_talk_v2.csv": 10_000, |
| "wizard_history_v2.csv": 10_000, |
| "geography_v2.csv": 100_000, |
| "biology_v2.csv": 100_000, |
| "social_etiquette_v2.csv": 10_000, |
| "history_cz_wizard_v2.csv": 100_000, |
| "cooking_v2.csv": 100_000, |
| "creative/stories.csv": 25_000, |
| "creative/essays_and_slohy.csv": 25_000, |
| "creative/poetry.csv": 20_000, |
| "creative/songs.csv": 20_000, |
| "creative/rhyme_craft.csv": 10_000, |
| "benchmarks/logic.csv": 10_000, |
| "benchmarks/math.csv": 10_000, |
| "benchmarks/instruction_following.csv": 10_000, |
| "benchmarks/truthfulness.csv": 10_000, |
| "benchmarks/context_reading.csv": 10_000, |
| "benchmarks/temporal_spatial.csv": 10_000, |
| "benchmarks/causal_counterfactual.csv": 10_000, |
| "benchmarks/language.csv": 10_000, |
| "benchmarks/planning_tools.csv": 10_000, |
| "benchmarks/code_data.csv": 10_000, |
| } |
| TAGS = { |
| "<|im_start|>system": 1, |
| "<|im_start|>user": 1, |
| "<|im_start|>assistant": 1, |
| "<thought>": 1, |
| "</thought>": 1, |
| "<|im_end|>": 3, |
| } |
|
|
|
|
| def last_word(line: str) -> str: |
| line = line.casefold().strip() |
| line = re.sub(r"[^a-záčďéěíňóřšťúůýž]+$", "", line) |
| return line.split()[-1] |
|
|
|
|
| def answer_lines(text: str) -> list[str]: |
| answer = text.split("</thought>\n", 1)[1].rsplit("<|im_end|>", 1)[0] |
| return answer.splitlines() |
|
|
|
|
| def validate_file(rel: str, expected: int, global_hashes: set[bytes]) -> dict: |
| path = OUT / rel |
| if not path.exists(): |
| raise AssertionError(f"Missing {rel}") |
| count = 0 |
| unique = set() |
| min_len = 10**9 |
| max_len = 0 |
| total_len = 0 |
| forbidden = 0 |
| tag_errors = 0 |
| with path.open("r", encoding="utf-8", newline="") as f: |
| reader = csv.DictReader(f) |
| if reader.fieldnames != ["text"]: |
| raise AssertionError(f"{rel}: expected one text column, got {reader.fieldnames}") |
| for row in reader: |
| count += 1 |
| text = row["text"] |
| digest = hashlib.blake2b(text.encode("utf-8"), digest_size=16).digest() |
| if digest in unique: |
| raise AssertionError(f"{rel}: duplicate row {count}") |
| unique.add(digest) |
| |
| |
| if digest in global_hashes: |
| raise AssertionError(f"cross-file duplicate at {rel}:{count}") |
| global_hashes.add(digest) |
| for tag, n in TAGS.items(): |
| if text.count(tag) != n: |
| tag_errors += 1 |
| raise AssertionError(f"{rel}:{count}: malformed ChatML tag {tag}") |
| if "\x00" in text: |
| raise AssertionError(f"{rel}:{count}: NUL byte") |
| if rel in {"wizard_history_v2.csv", "history_cz_wizard_v2.csv"} and "magi" in text.casefold(): |
| forbidden += 1 |
| raise AssertionError(f"{rel}:{count}: forbidden word root") |
| n = len(text) |
| min_len = min(min_len, n); max_len = max(max_len, n); total_len += n |
| if count != expected: |
| raise AssertionError(f"{rel}: expected {expected}, got {count}") |
| return { |
| "path": f"data/additions/{rel}", |
| "rows": count, |
| "bytes": path.stat().st_size, |
| "sha256": text_sha256(path), |
| "unique_rows": len(unique), |
| "min_chars": min_len, |
| "max_chars": max_len, |
| "avg_chars": round(total_len / count, 2), |
| "tag_errors": tag_errors, |
| "forbidden_hits": forbidden, |
| } |
|
|
|
|
| def validate_rhymes() -> dict: |
| checks = {"poetry_rows": 0, "song_rows": 0, "poetry_bad": 0, "song_bad": 0} |
| p = OUT / "creative/poetry.csv" |
| with p.open(encoding="utf-8", newline="") as f: |
| for i, row in enumerate(csv.DictReader(f)): |
| lines = answer_lines(row["text"])[1:] |
| e, a, b, c, raw_d = digits(i, [2, 25, 10, 30, 30]) |
| d = (raw_d * 13 + a * 7 + b * 3 + c + 1) % 30 |
| pa, pb = RHYME_PAIRS[c], RHYME_PAIRS[d] |
| expected = [pa[0], pa[1], pb[0], pb[1]] if e == 0 else [pa[0], pb[0], pa[1], pb[1]] |
| checks["poetry_rows"] += 1 |
| if [last_word(x) for x in lines] != expected: |
| checks["poetry_bad"] += 1 |
| p = OUT / "creative/songs.csv" |
| with p.open(encoding="utf-8", newline="") as f: |
| for i, row in enumerate(csv.DictReader(f)): |
| lines = [x for x in answer_lines(row["text"]) if x and not x.startswith("**")] |
| e, a, b, c = digits(i, [10, 25, 10, 30]) |
| d = (a * 11 + b * 5 + c * 7 + e + 3) % 30 |
| pa, pb = RHYME_PAIRS[c], RHYME_PAIRS[d] |
| expected = [pa[0], pa[1], pb[0], pb[1], pb[0], pa[0], pb[1], pa[1]] |
| checks["song_rows"] += 1 |
| if [last_word(x) for x in lines] != expected: |
| checks["song_bad"] += 1 |
| if checks["poetry_bad"] or checks["song_bad"]: |
| raise AssertionError(f"Rhyme validation failed: {checks}") |
| return checks |
|
|
|
|
| def legacy_report() -> dict: |
| result = {} |
| for path in sorted(REPO.glob("*.csv")): |
| rows = 0; malformed = 0 |
| try: |
| with path.open(encoding="utf-8-sig", newline="") as f: |
| for row in csv.DictReader(f): |
| rows += 1 |
| text = row.get("text", "") |
| if any(text.count(tag) != n for tag, n in TAGS.items()): |
| malformed += 1 |
| except Exception as exc: |
| result[path.name] = {"error": repr(exc)} |
| continue |
| result[path.name] = {"rows": rows, "malformed_chatml_rows": malformed, "bytes": path.stat().st_size} |
| return result |
|
|
|
|
| def main() -> None: |
| global_hashes: set[bytes] = set() |
| files = [validate_file(rel, n, global_hashes) for rel, n in EXPECTED.items()] |
| rhyme = validate_rhymes() |
| additional_rows = sum(x["rows"] for x in files) |
| additional_bytes = sum(x["bytes"] for x in files) |
| manifest = { |
| "generated_at": "2026-07-30", |
| "repository": "nekam13/zbynka-dataset", |
| "language": "cs", |
| "schema": {"format": "CSV", "columns": ["text"], "conversation": "ChatML-like text with concise thought summary"}, |
| "purpose": "Additional thematic expansion; original repository files are retained.", |
| "additional_rows": additional_rows, |
| "additional_bytes": additional_bytes, |
| "additional_mib": round(additional_bytes / 2**20, 2), |
| "files": files, |
| "rhyme_validation": rhyme, |
| "global_exact_duplicate_count": 0, |
| "legacy_root_report": legacy_report(), |
| "assumptions": [ |
| "The word 'dalších' was interpreted literally: requested rows were added to, not substituted for, existing files.", |
| "Cooking was assigned 100,000 additional rows because the request asked to teach cooking but did not specify a count; this is explicitly documented.", |
| "Benchmark-like items are original synthetic tasks inspired by task families, not copied benchmark questions.", |
| "The hidden historical layer is intentionally in-universe and is marked separately from publicly documented history.", |
| ], |
| "quality_notes": [ |
| "Every new file has exact requested row count and exact-row uniqueness, including across new files.", |
| "Wizard and integrated-history additions contain no case-insensitive 'magi' fragment.", |
| "Poetry and song rhyme schemes were checked row by row (40,000 rows).", |
| "Programmatic checks do not replace human review of factual, stylistic, or safety-sensitive content.", |
| ], |
| "benchmark_inspiration": [ |
| "https://github.com/google/BIG-bench", |
| "https://github.com/openai/simple-evals", |
| "https://aclanthology.org/2025.tacl-1.50/", |
| "https://github.com/MFajcik/benczechmark-leaderboard", |
| "https://github.com/simecek/MiniCzechBenchmark", |
| ], |
| "hub_upload_docs": [ |
| "https://huggingface.co/docs/huggingface_hub/guides/upload", |
| "https://huggingface.co/docs/huggingface_hub/package_reference/hf_api", |
| ], |
| } |
| out = REPO / "generation" / "manifest.json" |
| out.write_text(json.dumps(manifest, ensure_ascii=False, indent=2) + "\n", encoding="utf-8") |
| print(json.dumps({"additional_rows": additional_rows, "additional_mib": manifest["additional_mib"], "files": len(files), "rhyme": rhyme}, ensure_ascii=False)) |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
