fixtures/fetch_fixtures.py

#!/usr/bin/env python3
"""Build multilingual + modality text fixtures for tokenizer benchmarks.

Streams a few MB of real text per language/modality from public Hugging Face
datasets and writes plain .txt files into lang/ and modalities/ next to this
script, plus a fixtures_manifest.json recording source dataset, config,
split, pinned revision SHA, and sizes. See FIXTURES.md for the
human-readable version.

Usage:
    uv run --python 3.12 --with 'datasets>=3.2' fetch_fixtures.py [name ...]

With no args, builds every fixture. Names are file stems (e.g. hin_Deva).

Safety: HF_DATASETS_TRUST_REMOTE_CODE is forced off — only parquet-native
datasets are read, no dataset loading scripts ever execute. Downloaded
content is treated strictly as text data.
"""
from __future__ import annotations

import json
import os
import sys
import time
from pathlib import Path

os.environ["HF_DATASETS_TRUST_REMOTE_CODE"] = "0"

from datasets import load_dataset  # noqa: E402
from huggingface_hub import HfApi  # noqa: E402

HERE = Path(__file__).resolve().parent
MANIFEST = HERE / "fixtures_manifest.json"
MB = 1024 * 1024
LANG_TARGET = 5 * MB

FINEWEB2 = "HuggingFaceFW/fineweb-2"
FINEWEB2_CONFIGS = [
    "hin_Deva",  # Hindi, Devanagari
    "ben_Beng",  # Bengali
    "tam_Taml",  # Tamil
    "rus_Cyrl",  # Russian, Cyrillic
    "arb_Arab",  # Arabic (MSA), RTL
    "heb_Hebr",  # Hebrew, RTL
    "jpn_Jpan",  # Japanese, mixed kanji/kana
    "kor_Hang",  # Korean, Hangul
    "cmn_Hani",  # Mandarin, Han
    "tha_Thai",  # Thai, no word separators
    "ell_Grek",  # Greek
    "kat_Geor",  # Georgian
    "amh_Ethi",  # Amharic, Ethiopic
]

# the-stack-smol would be preferable but is gated; these are pinned-SHA GitHub
# tarballs instead, read in memory only (never extracted to disk).
CODE_REPOS = [
    ("python", "django/django", (".py",)),
    ("javascript", "jquery/jquery", (".js",)),
    ("typescript", "vuejs/core", (".ts",)),
    ("rust", "tokio-rs/tokio", (".rs",)),
    ("go", "caddyserver/caddy", (".go",)),
    ("java", "junit-team/junit5", (".java",)),
    ("c", "redis/redis", (".c", ".h")),
    ("cpp", "fmtlib/fmt", (".cc", ".h")),
]
CODE_SKIP_DIRS = ("deps/", "vendor/", "node_modules/", "third_party", "dist/")

_api = HfApi()
_shas: dict[str, str] = {}


def pinned_sha(repo: str) -> str:
    if repo not in _shas:
        _shas[repo] = _api.dataset_info(repo).sha
    return _shas[repo]


def stream(repo: str, split: str = "train", **kw):
    return load_dataset(repo, split=split, streaming=True, revision=pinned_sha(repo), **kw)


def fineweb2_docs(cfg: str):
    try:
        ds, split = stream(FINEWEB2, split="test", name=cfg), "test"
    except Exception:
        ds, split = stream(FINEWEB2, split="train", name=cfg), "train"
    return (ex["text"] for ex in ds), {"config": cfg, "split": split}


def eng_docs():
    ds = stream("HuggingFaceFW/fineweb", name="sample-10BT")
    return (ex["text"] for ex in ds), {"config": "sample-10BT", "split": "train"}


def code_docs():
    import tarfile
    import urllib.request

    per_lang = (6 * MB) // len(CODE_REPOS)
    pins: dict[str, str] = {}

    def head_sha(repo: str) -> str:
        req = urllib.request.Request(
            f"https://api.github.com/repos/{repo}/commits/HEAD",
            headers={"Accept": "application/vnd.github.sha"},
        )
        with urllib.request.urlopen(req) as r:
            return r.read().decode().strip()

    def gen():
        for lang, repo, exts in CODE_REPOS:
            sha = head_sha(repo)
            pins[repo] = sha
            url = f"https://codeload.github.com/{repo}/tar.gz/{sha}"
            got = 0
            with urllib.request.urlopen(url) as r:
                with tarfile.open(fileobj=r, mode="r|gz") as tar:
                    for member in tar:
                        if got >= per_lang:
                            break
                        path = member.name.lower()
                        if not member.isfile() or member.size > 200_000:
                            continue
                        if not path.endswith(exts) or any(d in path for d in CODE_SKIP_DIRS):
                            continue
                        text = tar.extractfile(member).read().decode("utf-8", errors="replace")
                        yield f"// {member.name}\n{text}"
                        got += len(text.encode("utf-8"))

    # pins is filled while gen() runs; build() reads meta only after consuming docs
    return gen(), {"revision": pins}


def math_docs():
    ds = stream("open-web-math/open-web-math")
    return (ex["text"] for ex in ds), {"split": "train"}


def agentic_docs():
    # single-shard hf_hub_download instead of streaming: the 249MB shard's
    # range-reads time out on unauthenticated connections
    import pyarrow.parquet as pq
    from huggingface_hub import hf_hub_download

    repo = "SWE-bench/SWE-smith-trajectories"
    shard = "data/tool-00000-of-00008.parquet"
    path = hf_hub_download(repo, shard, repo_type="dataset", revision=pinned_sha(repo))

    def rows():
        pf = pq.ParquetFile(path)
        for batch in pf.iter_batches(columns=["messages"], batch_size=8):
            yield from batch.column("messages").to_pylist()

    def render(content):
        if isinstance(content, str):
            return content
        if isinstance(content, list):
            return "\n".join(
                b.get("text", json.dumps(b, ensure_ascii=False)) if isinstance(b, dict) else str(b)
                for b in content
            )
        return json.dumps(content, ensure_ascii=False)

    def gen():
        for raw in rows():
            parts = []
            for msg in json.loads(raw):
                block = f"[{msg.get('role', '?')}]\n{render(msg.get('content', ''))}"
                if msg.get("tool_calls"):
                    block += "\n" + json.dumps(msg["tool_calls"], ensure_ascii=False)
                parts.append(block)
            yield "\n\n".join(parts)

    return gen(), {"split": "tool", "config": shard}


FIXTURES: dict[str, dict] = {
    cfg: {"source": FINEWEB2, "docs": (lambda c=cfg: fineweb2_docs(c)), "target": LANG_TARGET, "dir": "lang"}
    for cfg in FINEWEB2_CONFIGS
}
FIXTURES["eng_Latn"] = {"source": "HuggingFaceFW/fineweb", "docs": eng_docs, "target": LANG_TARGET, "dir": "lang"}
FIXTURES["code_mixed"] = {"source": "github.com (pinned repo tarballs)", "docs": code_docs, "target": 6 * MB, "dir": "modalities"}
FIXTURES["math_latex"] = {"source": "open-web-math/open-web-math", "docs": math_docs, "target": LANG_TARGET, "dir": "modalities"}
FIXTURES["agentic_swe"] = {"source": "SWE-bench/SWE-smith-trajectories", "docs": agentic_docs, "target": LANG_TARGET, "dir": "modalities"}


def build(name: str, spec: dict) -> dict:
    docs, meta = spec["docs"]()
    out = HERE / spec["dir"] / f"{name}.txt"
    out.parent.mkdir(exist_ok=True)
    written = n_docs = 0
    with open(out, "w", encoding="utf-8") as f:
        for text in docs:
            if not text or not text.strip():
                continue
            chunk = text.replace("\0", "").rstrip("\n") + "\n\n"
            f.write(chunk)
            written += len(chunk.encode("utf-8"))
            n_docs += 1
            if written >= spec["target"]:
                break
    if "revision" not in meta:
        meta["revision"] = pinned_sha(spec["source"])
    return {
        "file": f"{spec['dir']}/{out.name}",
        "source": spec["source"],
        **meta,
        "bytes": written,
        "docs": n_docs,
        "built_at": time.strftime("%Y-%m-%d", time.gmtime()),
        "builder": Path(__file__).name,
    }


def main() -> int:
    names = sys.argv[1:] or list(FIXTURES)
    unknown = [n for n in names if n not in FIXTURES]
    if unknown:
        print(f"unknown fixtures: {unknown}\navailable: {list(FIXTURES)}", file=sys.stderr)
        return 2

    manifest = json.loads(MANIFEST.read_text()) if MANIFEST.exists() else {}
    failures = []
    for name in names:
        print(f"building {name} ...", flush=True)
        try:
            entry = build(name, FIXTURES[name])
        except Exception as e:
            print(f"  FAILED {name}: {type(e).__name__}: {e}", file=sys.stderr, flush=True)
            failures.append(name)
            continue
        manifest[name] = entry
        MANIFEST.write_text(json.dumps(manifest, indent=2, sort_keys=True) + "\n")
        print(f"  {entry['bytes'] / MB:.1f} MB, {entry['docs']} docs", flush=True)

    if failures:
        print(f"failed: {failures}", file=sys.stderr)
    return 1 if failures else 0


if __name__ == "__main__":
    sys.exit(main())