Safetensors
English
duplex-0 / scripts /duplex /model /hf_release.py
RonanMcGovern's picture
Data step builds from the public datasets
f686fdd verified
Raw History Blame Contribute Delete
22.8 kB
"""Verify the published duplex token datasets and write one trainer manifest per stage.
Reads the private Hugging Face layout (`manifest.json`, `<run>/train.parquet`, `<run>/validation.parquet`, `<run>/test.parquet`, `<run>/report.json`,
`<run>/tokens.npz`) downloaded at pinned revisions. Checks every file against `manifest.json`, every
token array against its row (member hash, dtype, `[8, frames]` shape, code range, frame count from the
24 kHz sample count) and each separated speaker track the same way. Each array is extracted byte for byte
to `<out>/tokens/<key>.npy`, so its hash stays the row's, and the stage manifests use the trainer's
schema-version-2 rows (`tokens.py`): two audio tracks and two text tracks per span, assigned as training
readiness describes. Each text track is the words starting in the span's complete frames, tokenized with
Moshi's SentencePiece model, as int32 `(frame, id)` rows in `<out>/text/`.
A row's published `split` field decides its split. The data owner gives a group one split in every
dataset, so a span with different splits in different datasets fails the build. Revisions without it hold out one in 20 groups by a
salted hash (a group is the podcast feed or the YODAS recording), the same groups in every stage, so the
nested sets (conversational within mid-training within podcast pretraining) cannot leak. Every stage
manifest holds every validation span, tagged in `validation_sets` with the stages whose datasets hold it.
`--stages` builds a subset of the stages and reads only their datasets; a stage left out is absent from
`validation_sets` and from the nesting check.
"""
import argparse
import hashlib
import io
import json
import math
import zipfile
from pathlib import Path
import numpy as np
CODEC = {
"repo": "kyutai/mimi",
"sample_rate": 24000,
"codebooks": 8,
"cardinality": 2048,
"fps": 12.5,
}
HOP = 1920
TOKENIZER = (
"kyutai/moshiko-pytorch-bf16",
"tokenizer_spm_32k_3.model",
"2bfc9ae6e89079a5cc7ed2a68436010d91a3d289",
)
STAGES = {
"pretraining": ("duplex-pre-training-yodas3-mimi", "duplex-pre-training-podcasts-mimi"),
"mid_training": ("duplex-mid-training-podcasts-mimi",),
"conversational": ("duplex-conversational-podcasts-mimi",),
}
def digest(data):
return hashlib.sha256(data).hexdigest()
def group(row):
"""Podcast feeds share hosts across episodes, so a feed is the unit held out; YODAS has no uploader."""
return row.get("split_group") or row["source"].get("rss_url") or row["recording_id"]
def held_out(name, salt, every):
return int(digest(f"{salt}:{name}".encode()), 16) % every == 0
def check_array(data, record, where, problems):
"""Returns the array when its bytes, dtype, shape, range and frame count match the record."""
if digest(data) != record["sha256"]:
problems.append({**where, "problem": "token hash mismatch"})
return None
array = np.load(io.BytesIO(data), allow_pickle=False)
# Independent 60 s pieces (rows before 14 November) and pieces with dropped context both
# join to one frame per started 1,920 samples.
expected = -(-record["samples_24k"] // HOP)
if array.dtype != np.uint16 or array.shape != (8, record["frames"]):
problems.append({**where, "problem": f"array {array.dtype} {array.shape}"})
return None
if record["frames"] != expected:
problems.append(
{**where, "problem": f"frames {record['frames']} != {expected}"}
)
return None
if array.size and int(array.max()) >= 2048:
problems.append({**where, "problem": "code out of range"})
return None
return array
def row_files(folder):
"""A run folder's row files: Parquet per split, or the JSON Lines layouts of revisions before 17 November."""
for suffix in ("parquet", "jsonl"):
if found := [p for p in (folder / f"{name}.{suffix}" for name in ("train", "validation", "test")) if p.exists()]:
return found
return [folder / "spans.jsonl"]
def read_rows(file):
"""Yields a row file's rows as dicts, a few at a time; a Parquet row holds null where a JSON row had no key."""
if file.suffix == ".parquet":
import pyarrow.parquet as pq
for batch in pq.ParquetFile(file).iter_batches(batch_size=50):
yield from batch.to_pylist()
else:
with file.open() as lines:
yield from (json.loads(line) for line in lines if line.strip())
def audit_repo(path):
"""Checks a repository and returns (manifest, rows, problems, totals). rows yields
(run, row) pairs one run at a time, so a repository need not fit in memory;
problems and totals are complete once rows is exhausted, and a row's own
problems are recorded before it is yielded."""
manifest = json.loads((path / "manifest.json").read_text())
problems, totals = [], {}
def rows():
for name, info in manifest["files"].items():
data = (path / name).read_bytes()
if len(data) != info["bytes"] or digest(data) != info["sha256"]:
problems.append({"file": name, "problem": "file hash or size mismatch"})
for run, expected in manifest["runs"].items():
# A run with no spans has no Parquet files; a run that expects spans must have its row file.
spans = [
row
for file in row_files(path / run)
if expected["spans"] or file.exists()
for row in read_rows(file)
]
archive = zipfile.ZipFile(path / run / "tokens.npz")
members = {n.removesuffix(".npy") for n in archive.namelist()}
used = set()
hours = sum(s["seconds"] for s in spans) / 3600
stats = totals.setdefault(
run,
{
"spans": len(spans),
"hours": hours,
"manifest_spans": expected["spans"],
"manifest_hours": expected["hours"],
"stored_frames": 0,
"valid_frames": 0,
"track_arrays": 0,
"separation": {},
},
)
if (
len(spans) != expected["spans"]
or abs(hours - expected["hours"]) > 0.006
):
problems.append(
{"run": run, "problem": "span count or hours differ from manifest"}
)
for row in spans:
where = {
"run": run,
"recording_id": row["recording_id"],
"start_sample": row["start_sample"],
}
tokens = row["tokens"]
if tokens.get("status") != "ok" or tokens["key"] not in members:
problems.append(
{
**where,
"problem": f"tokens {tokens.get('status')}, array missing",
}
)
continue
data = archive.read(tokens["key"] + ".npy")
used.add(tokens["key"])
array = check_array(data, tokens, where, problems)
if array is None:
continue
rate = row["audio"]["sample_rate"]
span = row["end_sample"] - row["start_sample"]
if abs(tokens["samples_24k"] - span * 24000 / rate) > 1:
problems.append(
{**where, "problem": "24 kHz length disagrees with span"}
)
stats["stored_frames"] += tokens["frames"]
stats["valid_frames"] += tokens["samples_24k"] // HOP
separation = row.get("separation") or {"status": "not_run"}
status = separation["status"]
stats["separation"][status] = stats["separation"].get(status, 0) + 1
# Separated tracks of revisions before `tracks`, then the published tracks;
# a one-speaker row's track 1 is the mixed record, already checked.
published = [
t["tokens"]
for t in row.get("tracks") or []
if t["status"] == "ok" and t["tokens"]["key"] != tokens["key"]
]
listed = separation.get("tracks") or []
seen_keys = {t.get("key") for t in listed}
for track in listed + [
t for t in published if t["key"] not in seen_keys
]:
if track.get("status") != "ok" or track["key"] not in members:
problems.append(
{
**where,
"problem": f"track {track.get('status')}, array missing",
}
)
continue
used.add(track["key"])
stats["track_arrays"] += 1
if (
check_array(
archive.read(track["key"] + ".npy"), track, where, problems
)
is None
):
continue
if track["frames"] != tokens["frames"]:
problems.append(
{
**where,
"problem": "track frames differ from mixed stream",
}
)
yield run, row
if members - used:
problems.append(
{
"run": run,
"problem": f"{len(members - used)} unreferenced arrays",
}
)
return manifest, rows(), problems, totals
def text_rows(words, valid, tokenize):
"""(frame, id) rows of the words starting in the first `valid` frames, in word order."""
rows = []
for word in words:
text = word["word"].strip()
frame = math.floor(word["start_s"] * 12.5)
# The transcription service can time a punctuation mark on its own.
if not 0 <= frame < valid or not any(c.isalnum() for c in text):
continue
rows.extend((frame, i) for i in tokenize(text))
return np.array(rows, dtype=np.int32).reshape(-1, 2)
def tracks_of(row):
"""Audio token records and word lists of the span's two tracks.
A word list None is absent text; [] is text present with no words. From the
published `tracks` field when the revision has it, else from `separation`.
"""
words = row.get("words") if row.get("transcription_status") == "ok" else None
published = row.get("tracks")
if published is not None:
audio, text = [], []
for track in published:
audio.append(track["tokens"] if track["status"] == "ok" else None)
if track["status"] == "empty":
text.append([])
elif track["status"] != "ok" or words is None or track["speaker"] is None:
text.append(None)
else:
text.append(track["words"])
# Spans with no published track, such as more than two diarized speakers,
# train as the mixed audio on track 1 without text.
if not any(audio):
return [row["tokens"], None], [None, None]
return audio, text
separation = row.get("separation") or {}
found = separation.get("tracks") or []
if (
separation.get("status") == "ok"
and len(found) == 2
and all(t.get("status") == "ok" for t in found)
):
assignment = separation.get("assignment") or {}
speakers = assignment.get("speakers") or []
if words is None or not assignment.get("resolved") or len(speakers) != 2:
return found, [None, None]
return found, [[w for w in words if w["speaker"] == s] for s in speakers]
speakers = (row.get("stats") or {}).get("speakers")
if words is not None and speakers == 1:
return [row["tokens"], None], [words, []]
return [row["tokens"], None], [None, None]
def valid_frames(row):
"""Complete frames inside the source window on the source clock, as the loader requires."""
rate = row["audio"]["sample_rate"]
valid = row["tokens"]["samples_24k"] // HOP
while valid * rate * 2 > (row["end_sample"] - row["start_sample"]) * 25:
valid -= 1
return valid
def trainer_row(row, run, repo, split, revision, tokenize, out):
tokens = row["tokens"]
valid = valid_frames(row)
audio, words = tracks_of(row)
text = []
for index, track in enumerate(words, 1):
if track is None:
text.append(None)
continue
array = text_rows(track, valid, tokenize)
if not len(array):
text.append({"count": 0})
continue
path = Path("text") / f"{tokens['key']}-text{index}.npy"
buffer = io.BytesIO()
np.save(buffer, array, allow_pickle=False)
(out / path).write_bytes(buffer.getvalue())
text.append(
{
"path": str(path),
"sha256": digest(buffer.getvalue()),
"count": len(array),
}
)
return {
"schema_version": 2,
"recording_id": row["recording_id"],
"source_group": group(row),
"split": split,
"split_status": "published_group_split"
if "split" in row
else "provisional_feed_and_recording_holdout",
"source": row["source"],
"audio": row["audio"],
"metadata": {
# Revisions published before 17 November carry these; later ones keep them out of the rows.
"policy": row.get("policy"),
"qualification_status": row.get("qualification_status"),
},
"window": {
"start_sample": row["start_sample"],
"end_sample": row["end_sample"],
},
"codec": {
**CODEC,
"revision": revision,
"implementation": tokens["codec"]["codec_implementation"],
},
"tokens": {
"frames": tokens["frames"],
"valid_frames": valid,
"dtype": "uint16",
"piece_frame_boundaries": tokens.get("piece_frame_boundaries") or [],
"mixed_sha256": tokens["sha256"],
"tracks": [
None
if t is None
else {"path": f"tokens/{t['key']}.npy", "sha256": t["sha256"]}
for t in audio
],
},
"text": text,
"span": {
"repo": repo,
"run": run,
"key": tokens["key"],
"dataset": row["dataset"],
},
}
def check_splits(staged):
"""Fail when one span has different splits in different datasets: a stage would
train on a span another stage holds out."""
splits = {}
for stage_rows in staged.values():
for identity, row in stage_rows.items():
splits.setdefault(identity, set()).add(row["split"])
mixed = {k: sorted(v) for k, v in splits.items() if len(v) > 1}
if mixed:
examples = list(mixed.items())[:5]
raise ValueError(
f"{len(mixed)} spans have different splits across datasets: {examples}"
)
def main():
p = argparse.ArgumentParser(description=__doc__)
p.add_argument(
"--raw", type=Path, required=True, help="Directory holding the downloaded repos"
)
p.add_argument("--out", type=Path, required=True)
p.add_argument("--salt", default="nano-s2s-sanity-v1")
p.add_argument(
"--every", type=int, default=20, help="Hold out one group in this many"
)
p.add_argument(
"--tokenizer", type=Path, help="Moshi's SentencePiece model; default: download"
)
p.add_argument(
"--stages",
nargs="+",
choices=list(STAGES),
default=list(STAGES),
help="Stages to build; each needs its datasets under --raw. Default: all",
)
args = p.parse_args()
stages = {stage: STAGES[stage] for stage in STAGES if stage in args.stages}
import sentencepiece
tokenizer = args.tokenizer
if tokenizer is None:
from huggingface_hub import hf_hub_download
tokenizer = hf_hub_download(TOKENIZER[0], TOKENIZER[1], revision=TOKENIZER[2])
model = sentencepiece.SentencePieceProcessor(model_file=str(tokenizer))
tokenize = model.encode
for folder in ("tokens", "text"):
(args.out / folder).mkdir(parents=True, exist_ok=True)
audit, by_span, revisions, sources = {}, {}, set(), {}
staged = {stage: {} for stage in stages}
for stage, repos in stages.items():
for repo in repos:
_, rows, problems, totals = audit_repo(args.raw / repo)
audit[repo] = {"problems": problems, "runs": totals}
archives = {}
for run, row in rows:
if any(
p.get("recording_id") == row["recording_id"]
and p.get("start_sample") == row["start_sample"]
for p in problems
):
continue
revisions.add(row["tokens"]["codec"]["codec_revision"])
identity = (row["recording_id"], row["start_sample"], row["end_sample"])
seen = by_span.setdefault(identity, (row["tokens"]["sha256"], set()))
if seen[0] != row["tokens"]["sha256"]:
problems.append(
{
"run": run,
"recording_id": row["recording_id"],
"problem": "same span has different tokens across repos",
}
)
seen[1].add(repo)
sources.setdefault(identity, set()).add(stage)
archive = archives.get(run) or zipfile.ZipFile(
args.raw / repo / run / "tokens.npz"
)
archives[run] = archive
audio, _ = tracks_of(row)
for record in [row["tokens"], *audio]:
if record is None:
continue
target = args.out / "tokens" / f"{record['key']}.npy"
if not target.exists():
target.write_bytes(archive.read(record["key"] + ".npy"))
if "split" in row:
split = row["split"]
else:
split = (
"validation"
if held_out(group(row), args.salt, args.every)
else "train"
)
staged[stage][identity] = trainer_row(
row, run, repo, split, None, tokenize, args.out
)
check_splits(staged)
if len(revisions) != 1:
raise ValueError(f"Expected one codec revision, found {sorted(revisions)}")
(revision,) = revisions
# Nesting: each conversational span is a mid-training span and each mid-training span a podcast pretraining span.
nesting = {}
for inner, outer in (
("duplex-conversational-podcasts-mimi", "duplex-mid-training-podcasts-mimi"),
("duplex-mid-training-podcasts-mimi", "duplex-pre-training-podcasts-mimi"),
):
if not {inner, outer} <= {repo for repos in stages.values() for repo in repos}:
continue
inside = [repos for _, repos in by_span.values() if inner in repos]
nesting[f"{inner} in {outer}"] = {
"spans": len(inside),
"missing": sum(outer not in repos for repos in inside),
}
# Every stage scores every validation set: its manifest holds every validation span.
held = {}
for stage in stages:
for identity, row in staged[stage].items():
if row["split"] == "validation":
held.setdefault(identity, row)
for identity, row in held.items():
row["validation_sets"] = sorted(sources[identity])
summary = {
"salt": args.salt,
"every": args.every,
"codec_revision": revision,
"tokenizer": {
"repo": TOKENIZER[0],
"file": TOKENIZER[1],
"revision": TOKENIZER[2],
},
"audit": audit,
"nesting": nesting,
"stages": {},
}
for stage in stages:
rows = [r for r in staged[stage].values() if r["split"] != "validation"]
rows += list(held.values())
lines = []
for row in rows:
row["codec"]["revision"] = revision
lines.append(json.dumps(row, sort_keys=True) + "\n")
path = args.out / f"{stage}.jsonl"
path.write_text("".join(lines))
split = {}
for row in rows:
name = row["split"]
if name == "validation":
name = "validation:" + ",".join(row["validation_sets"])
entry = split.setdefault(
name,
{
"spans": 0,
"hours": 0.0,
"groups": set(),
"recordings": set(),
"two_tracks": 0,
"text_tracks": 0,
"text_tokens": 0,
},
)
entry["spans"] += 1
entry["hours"] += row["tokens"]["valid_frames"] / 12.5 / 3600
entry["groups"].add(row["source_group"])
entry["recordings"].add(row["recording_id"])
entry["two_tracks"] += all(row["tokens"]["tracks"])
entry["text_tracks"] += sum(t is not None for t in row["text"])
entry["text_tokens"] += sum(t["count"] for t in row["text"] if t)
summary["stages"][stage] = {
"manifest": path.name,
"manifest_sha256": digest(path.read_bytes()),
"splits": {
k: {**v, "groups": len(v["groups"]), "recordings": len(v["recordings"])}
for k, v in sorted(split.items())
},
}
(args.out / "summary.json").write_text(json.dumps(summary, indent=1) + "\n")
print(
json.dumps(
{
"stages": summary["stages"],
"nesting": nesting,
"problems": {r: a["problems"][:20] for r, a in audit.items()},
"problem_counts": {r: len(a["problems"]) for r, a in audit.items()},
},
indent=1,
)
)
if __name__ == "__main__":
main()