xscript-eval / src /xscript /data /fineweb.py
jvonrad's picture
Upload src/xscript/data/fineweb.py with huggingface_hub
694072a verified
Raw
History Blame Contribute Delete
14.9 kB
"""Model-training text pools from FineWeb-HQ (EN) / FineWeb2-HQ (DE/FR/AR/ZH).
Both datasets come from the same lab and the same XLM-R-embedding quality
classifier (top-10%% selection), keeping filtering methodology constant across
all five languages -- with two documented exceptions: FineWeb2-HQ's arb_Arab
split is only ~29GB total (85 parquet files, confirmed exhausted 2026-07-13),
and fra_Latn -HQ fully exhausted at 124.1GB (2026-07-14), both far short of
what our token budgets need. `FALLBACK_SOURCES` lets a language keep pulling
from a second, less-strict source (FineWeb2's own dedup+quality-filtered
"train" split, one filter step short of the top-decile HQ cut -- still same
lab/pipeline) once its primary source is exhausted, rather than epoch heavily
over a small pool. This is a real, deliberate deviation from strict
filter-parity for these two languages; note it in the thesis.
We stream parquet with column pruning (only `text`), so the large `embeddings`
column in FineWeb2-HQ is never downloaded. Output pools are zstd jsonl shards
of ~1GB uncompressed text. The first parquet file of the primary source is
reserved exclusively for the in-domain eval holdout (never enters the pool).
"""
import json
from pathlib import Path
from ..langs import LANGS
from ..paths import MANIFEST_CACHE, POOLS, HOLDOUT, pool_dir, ensure
POOL_SHARD_BYTES = 1 << 30 # uncompressed text per pool shard
HOLDOUT_BYTES = 30 * (1 << 20) # 30MB per language
FALLBACK_SOURCES: dict[str, tuple[str, str]] = {
"ar": ("HuggingFaceFW/fineweb-2", "data/arb_Arab/train"),
# fra_Latn -HQ (epfml/FineWeb2-HQ) fully exhausted 2026-07-14 at 124.1GB,
# short of the 152.2GB budget -- same fallback pattern as ar.
"fr": ("HuggingFaceFW/fineweb-2", "data/fra_Latn/train"),
}
def _sources_for(lang: str) -> list[tuple[str, str]]:
"""[(repo, subdir), ...] in priority order: primary (quality-filtered -HQ)
first, then any FALLBACK_SOURCES entry once the primary is exhausted."""
L = LANGS[lang]
srcs = [(L.fineweb_repo, L.fineweb_subdir)]
if lang in FALLBACK_SOURCES:
srcs.append(FALLBACK_SOURCES[lang])
return srcs
def _list_parquets(repo: str, subdir: str) -> list[str]:
"""Manifest of parquet files under repo/subdir, round-robin across CC dumps
for raw FineWeb-HQ (EN)."""
from huggingface_hub import HfApi
cache = ensure(MANIFEST_CACHE / "fineweb_hq_pools") / \
f"{repo.replace('/', '__')}__{subdir.replace('/', '_')}.json"
if cache.exists():
return json.loads(cache.read_text())
api = HfApi()
files = [e.path for e in api.list_repo_tree(repo, subdir,
repo_type="dataset", recursive=True)
if e.__class__.__name__ == "RepoFile" and e.path.endswith(".parquet")]
if repo == "epfml/FineWeb-HQ":
# data/CC-MAIN-YYYY-WW/000_xxxxx.parquet: interleave dumps so the pool
# spans the full 2013-2024 crawl range like FineWeb2-HQ does.
by_dump: dict[str, list[str]] = {}
for f in sorted(files):
by_dump.setdefault(f.split("/")[1], []).append(f)
out = []
for i in range(max(len(v) for v in by_dump.values())):
for d in sorted(by_dump):
if i < len(by_dump[d]):
out.append(by_dump[d][i])
files = out
else:
files = sorted(files)
cache.write_text(json.dumps(files))
return files
def _iter_texts(repo: str, path_in_repo: str):
import pyarrow.parquet as pq
from huggingface_hub import HfFileSystem
fs = HfFileSystem()
with fs.open(f"datasets/{repo}/{path_in_repo}", "rb") as f:
pf = pq.ParquetFile(f)
for rg in range(pf.num_row_groups):
tbl = pf.read_row_group(rg, columns=["text"])
for t in tbl.column("text").to_pylist():
if t:
yield t
class _PoolWriter:
def __init__(self, out_dir: Path, prefix: str = "pool", start_idx: int = -1,
total_bytes: int = 0, total_docs: int = 0):
import zstandard
self.dir = ensure(out_dir)
self.prefix = prefix
self.zstd = zstandard
self.idx = start_idx
self.cur = None
self.cur_bytes = 0
self.total_bytes = total_bytes
self.total_docs = total_docs
self._roll()
def _roll(self):
if self.cur:
self.cur.close()
self.idx += 1
self.cur_bytes = 0
raw = open(self.dir / f"{self.prefix}_{self.idx:05d}.jsonl.zst", "wb")
self.cur = self.zstd.ZstdCompressor(level=3).stream_writer(raw, closefd=True)
def write(self, text: str):
line = json.dumps({"text": text}, ensure_ascii=False) + "\n"
b = line.encode("utf-8")
self.cur.write(b)
n = len(text.encode("utf-8"))
self.cur_bytes += n
self.total_bytes += n
self.total_docs += 1
if self.cur_bytes >= POOL_SHARD_BYTES:
self._roll()
def close(self):
if self.cur:
self.cur.close()
self.cur = None
CHECKPOINT_EVERY_N_FILES = 3 # bound on re-downloaded work if the process dies
def _next_shard_idx(out_dir: Path, prefix: str = "pool") -> int:
existing = sorted(out_dir.glob(f"{prefix}_*.jsonl.zst"))
if not existing:
return -1
stem = existing[-1].name[len(prefix) + 1:] # "NNNNN.jsonl.zst"
return int(stem.split(".", 1)[0])
def build_pool(lang: str, budget_bytes: float, holdout_bytes: int = HOLDOUT_BYTES) -> dict:
"""Build (or resume) a language's text pool.
Crash-resumable: every CHECKPOINT_EVERY_N_FILES source files, the current
shard is closed (a truncated zstd frame from a mid-write kill can't be
decoded, so a checkpoint never leaves a shard half-written) and
`stats.json` records which manifest files are already consumed plus the
index of that last cleanly-closed shard (`shard_idx`). A crash *between*
checkpoints (e.g. mid-way through one large parquet file) can still leave
a higher-numbered shard on disk with an unterminated zstd frame -- since
none of its bytes were counted into `text_bytes`/`docs` at the last
checkpoint, resuming deletes any shard past `shard_idx` before writing
starts again, so a stray corrupt shard never lingers for `pack()` to trip
on. Re-running `xscript pool` after an interruption (SSH drop, node
hiccup, ^C, session teardown) picks up from there instead of
re-downloading from scratch. `files_consumed` entries are tagged
"repo::path" so a language with a FALLBACK_SOURCES entry can track
consumption across sources without collision; old untagged checkpoints
(single-source) are migrated on load by assuming the primary repo.
"""
sources = _sources_for(lang)
primary_repo, primary_subdir = sources[0]
first_files = _list_parquets(primary_repo, primary_subdir)
if not first_files:
raise RuntimeError(f"no parquet files found for {lang}")
out = pool_dir(lang)
stats_path = out / "stats.json"
resume = None
if stats_path.exists():
st = json.loads(stats_path.read_text())
if st["text_bytes"] >= budget_bytes * 0.99:
print(f"[pool] {lang}: cached ({st['text_bytes']/1e9:.1f}GB)")
return st
if st.get("files_consumed"):
resume = st
print(f"[pool] {lang}: resuming from checkpoint "
f"({st['text_bytes']/1e9:.1f}/{budget_bytes/1e9:.1f}GB, "
f"{len(st['files_consumed'])} files already consumed)")
if resume is None:
# holdout from the primary source's first file only; pool starts at the second
hw = _PoolWriter(HOLDOUT, prefix=lang)
got = 0
for t in _iter_texts(primary_repo, first_files[0]):
hw.write(t)
got += len(t.encode("utf-8"))
if got >= holdout_bytes:
break
hw.close()
used: list[str] = []
pw = _PoolWriter(out)
else:
got = resume["holdout_bytes"]
used = [u if "::" in u else f"{primary_repo}::{u}" for u in resume["files_consumed"]]
# shard_idx is missing on checkpoints written before this field existed;
# best-effort fall back to whatever's on disk (pre-existing behaviour).
last_good_idx = resume.get("shard_idx")
if last_good_idx is None:
last_good_idx = _next_shard_idx(out)
else:
for stray in out.glob("pool_*.jsonl.zst"):
idx = int(stray.name[len("pool_"):].split(".", 1)[0])
if idx > last_good_idx:
stray.unlink() # never checkpointed -- may be a truncated zstd frame
# The file *at* shard_idx is the one _checkpoint()'s _roll() had just
# opened (empty) when that checkpoint was written -- everything written
# into it since then, up to a crash, was never counted in text_bytes/
# docs. It's fine (already fully closed) if the run reached this point
# via the final `pw.close()` on graceful completion, but indistinguishable
# from a crash-mid-write from stats.json alone -- so verify by decoding.
at_idx = out / f"pool_{last_good_idx:05d}.jsonl.zst"
if at_idx.exists():
import io
import zstandard
try:
with open(at_idx, "rb") as raw:
reader = zstandard.ZstdDecompressor().stream_reader(raw)
for jline in io.TextIOWrapper(reader, encoding="utf-8"):
if jline.strip():
json.loads(jline)
except Exception as exc:
print(f"[pool] {lang}: {at_idx.name} failed validation ({exc}) "
f"-- discarding (never counted in checkpointed totals)")
at_idx.unlink()
pw = _PoolWriter(out, start_idx=last_good_idx,
total_bytes=resume["text_bytes"], total_docs=resume["docs"])
def _checkpoint():
pw._roll() # close the current shard so it's a complete, valid zstd frame
st = {"lang": lang, "budget_bytes": budget_bytes, "text_bytes": pw.total_bytes,
"docs": pw.total_docs, "holdout_bytes": got, "holdout_file": first_files[0],
"files_consumed": used, "shard_idx": pw.idx, "exhausted": False}
stats_path.write_text(json.dumps(st, indent=2))
done = False
for i, (repo, subdir) in enumerate(sources):
files = first_files if i == 0 else _list_parquets(repo, subdir)
pool_files = files[1:] if i == 0 else files # only the primary source reserves a holdout file
for f in pool_files:
tag = f"{repo}::{f}"
if tag in used:
continue
used.append(tag)
try:
for t in _iter_texts(repo, f):
pw.write(t)
if pw.total_bytes >= budget_bytes:
break
except Exception as exc:
print(f"[pool] WARN {tag}: {exc}")
used.pop() # not actually consumed -- retry it on the next run
if pw.total_bytes >= budget_bytes:
done = True
_checkpoint()
break
if len(used) % CHECKPOINT_EVERY_N_FILES == 0:
_checkpoint()
if len(used) % 20 == 0:
print(f"[pool] {lang}: {pw.total_bytes/1e9:.1f}/{budget_bytes/1e9:.1f}GB "
f"({len(used)} files, source {i+1}/{len(sources)}: {repo})")
if done:
break
if i + 1 < len(sources):
print(f"[pool] {lang}: source {i+1}/{len(sources)} ({repo}) exhausted at "
f"{pw.total_bytes/1e9:.1f}GB -> falling back to {sources[i+1][0]}")
pw.close()
st = {"lang": lang, "budget_bytes": budget_bytes, "text_bytes": pw.total_bytes,
"docs": pw.total_docs, "holdout_bytes": got, "holdout_file": first_files[0],
"files_consumed": used, "shard_idx": pw.idx,
"exhausted": pw.total_bytes < budget_bytes * 0.99}
stats_path.write_text(json.dumps(st, indent=2))
if st["exhausted"]:
print(f"[pool] WARNING {lang}: corpus exhausted at {pw.total_bytes/1e9:.1f}GB "
f"< budget {budget_bytes/1e9:.1f}GB -> training will epoch over this pool")
print(f"[pool] {lang}: {pw.total_bytes/1e9:.2f}GB text, {pw.total_docs} docs")
return st
def _measured_bytes_per_token(flavor: str = "unigram", condition: str = "destarved") -> dict[str, float]:
"""Real bytes/token per study language, measured on FLORES+ dev with the
tokenizer that will actually pack the pool text. Destarved is the more
byte-hungry of the two conditions (better fertility -> more input bytes
needed per token), so it's the binding case for sizing. Scripts vary a
lot (Arabic ~6.8 bytes/token vs Chinese ~4.0) -- returns {} (caller falls
back to a flat estimate) if the tokenizer or FLORES+ aren't ready yet.
"""
from .. import flores
from ..langs import tok_name
from ..paths import tokenizer_dir
from ..tok.wrapper import Tok
tdir = tokenizer_dir(tok_name(flavor, condition))
if not (tdir / "meta.json").exists():
return {}
try:
tok = Tok(tdir)
par = flores.load_parallel(list(LANGS), "dev")
except Exception as exc:
print(f"[pool] WARN: couldn't measure real bytes/token ({exc}); "
f"falling back to the flat estimate")
return {}
out = {}
for l, sents in par.items():
b = sum(len(s.encode("utf-8")) for s in sents)
t = sum(len(tok.encode(s)) for s in sents)
out[l] = b / t
return out
def plan_budgets(tokens_per_run: float = 30e9, est_bytes_per_token: float = 4.5,
safety: float = 1.15) -> dict[str, float]:
"""Per-language pool byte budgets.
Monolingual runs need the full token budget in one language; bilingual
runs need half. The pool must cover the *max* need across planned runs
under the worst-case (most byte-hungry, i.e. destarved) tokenizer.
Sized from each language's REAL measured bytes/token (destarved
tokenizer on FLORES+), not a flat guess -- a single constant badly
under/over-shoots per language given how much bytes/token varies by
script. `est_bytes_per_token` is only a fallback for languages the
measurement can't cover yet (e.g. before the tokenizer gate).
"""
need_tokens = {l: tokens_per_run for l in LANGS} # monolingual dominates
measured = _measured_bytes_per_token()
return {l: need_tokens[l] * measured.get(l, est_bytes_per_token) * safety
for l in LANGS}