flow-smart-segments-20261004 / code /src /sponsor_detection /data /smart_segment_annotations.py
CuriousDragon's picture
Publish verified Flow Smart Segments curated-category-encoder-ensemble-20261004
8fcf7c9 verified
Raw History Blame Contribute Delete
13.6 kB
from __future__ import annotations
import csv
import json
import os
import tempfile
import time
from collections import Counter, defaultdict
from dataclasses import asdict, dataclass
from datetime import UTC, datetime
from pathlib import Path
from typing import Callable, Sequence
from sponsor_detection.data.profile import (
REQUIRED_COLUMNS,
EligibilityPolicy,
eligibility_rejection_flags,
sha256_file,
write_json_atomic,
)
SMART_SEGMENT_ANNOTATION_SCHEMA_VERSION = 1
NORMALIZATION_POLICY_VERSION = "smart-segment-annotations/1"
DEDUPLICATION_POLICY_VERSION = "preserve/1"
SOURCE_NAME = "sponsorblock_mirror"
# Within-category normalization is opt-in. ``preserve`` keeps one canonical row
# per mirror annotation so raw counts stay reproducible; cross-category overlaps
# are always preserved regardless of policy.
DEDUPLICATION_POLICIES = ("preserve",)
def _parquet_schema(source: dict[str, object], policy: EligibilityPolicy, categories):
try:
import pyarrow as pa
except ImportError as error:
raise RuntimeError(
"Parquet extraction requires the data extra: pip install -e '.[data]'"
) from error
metadata = {
b"smart_segments.schema_version": str(
SMART_SEGMENT_ANNOTATION_SCHEMA_VERSION
).encode(),
b"smart_segments.normalization_policy": NORMALIZATION_POLICY_VERSION.encode(),
b"smart_segments.deduplication_policy": DEDUPLICATION_POLICY_VERSION.encode(),
b"smart_segments.categories": json.dumps(
sorted(categories), separators=(",", ":")
).encode(),
b"smart_segments.source_sha256": str(source.get("sha256") or "").encode(),
b"smart_segments.eligibility_policy": json.dumps(
asdict(policy), sort_keys=True, separators=(",", ":")
).encode(),
}
return pa.schema(
[
("video_id", pa.string()),
("segment_id", pa.string()),
("category", pa.string()),
("start_ms", pa.int64()),
("end_ms", pa.int64()),
("votes", pa.int32()),
("locked", pa.bool_()),
("incorrect_votes", pa.int32()),
("views", pa.int64()),
("time_submitted_ms", pa.int64()),
("video_duration_ms", pa.int64()),
("hidden", pa.bool_()),
("shadow_hidden", pa.bool_()),
("action_type", pa.string()),
("service", pa.string()),
("reputation", pa.float32()),
("is_eligible", pa.bool_()),
("rejection_flags", pa.list_(pa.string())),
("source", pa.string()),
],
metadata=metadata,
)
def _new_batch() -> dict[str, list[object]]:
return {
"video_id": [],
"segment_id": [],
"category": [],
"start_ms": [],
"end_ms": [],
"votes": [],
"locked": [],
"incorrect_votes": [],
"views": [],
"time_submitted_ms": [],
"video_duration_ms": [],
"hidden": [],
"shadow_hidden": [],
"action_type": [],
"service": [],
"reputation": [],
"is_eligible": [],
"rejection_flags": [],
"source": [],
}
def _append_annotation(
batch: dict[str, list[object]],
row: dict[str, str],
policy: EligibilityPolicy,
) -> bool:
start_time = float(row["startTime"])
end_time = float(row["endTime"])
video_duration = float(row["videoDuration"])
votes = int(row["votes"])
locked = int(row["locked"])
hidden = int(row["hidden"])
shadow_hidden = int(row["shadowHidden"])
flags = eligibility_rejection_flags(
policy,
service=row["service"],
action_type=row["actionType"],
start_time=start_time,
end_time=end_time,
votes=votes,
hidden=hidden,
shadow_hidden=shadow_hidden,
video_duration=video_duration,
)
batch["video_id"].append(row["videoID"])
batch["segment_id"].append(row["UUID"])
batch["category"].append(row["category"])
batch["start_ms"].append(round(start_time * 1000))
batch["end_ms"].append(round(end_time * 1000))
batch["votes"].append(votes)
batch["locked"].append(locked != 0)
batch["incorrect_votes"].append(int(row["incorrectVotes"]))
batch["views"].append(int(row["views"]))
batch["time_submitted_ms"].append(int(row["timeSubmitted"]))
batch["video_duration_ms"].append(
round(video_duration * 1000) if video_duration > 0 else None
)
batch["hidden"].append(hidden != 0)
batch["shadow_hidden"].append(shadow_hidden != 0)
batch["action_type"].append(row["actionType"])
batch["service"].append(row["service"])
batch["reputation"].append(float(row["reputation"]))
batch["is_eligible"].append(not flags)
batch["rejection_flags"].append(sorted(flags))
batch["source"].append(SOURCE_NAME)
return not flags
def build_smart_segment_annotations(
input_path: Path,
output_path: Path,
manifest_path: Path,
policy: EligibilityPolicy,
*,
categories: list[str],
deduplication_policy: str = "preserve",
batch_size: int = 100_000,
compression: str = "zstd",
compute_sha256: bool = True,
expected_source_sha256: str | None = None,
progress_every_rows: int = 0,
progress_callback: Callable[[int, int, float], None] | None = None,
) -> dict[str, object]:
"""Normalize the raw SponsorBlock mirror into the canonical annotation table.
Every audited row is retained with its category, timestamps, quality
metadata, eligibility verdict and provenance; ineligible rows stay in the
artifact with ``is_eligible = false``. Overlaps between different categories
are never collapsed.
"""
try:
import pyarrow as pa
import pyarrow.parquet as pq
except ImportError as error:
raise RuntimeError(
"Parquet extraction requires the data extra: pip install -e '.[data]'"
) from error
if batch_size <= 0:
raise ValueError("batch_size must be positive")
if deduplication_policy not in DEDUPLICATION_POLICIES:
raise ValueError(
f"unsupported deduplication policy: {deduplication_policy}; "
f"supported: {', '.join(DEDUPLICATION_POLICIES)}"
)
audited_categories = list(dict.fromkeys(categories))
if not audited_categories:
raise ValueError("at least one category must be audited")
category_set = set(audited_categories)
started = time.monotonic()
source_sha256 = sha256_file(input_path) if compute_sha256 else None
if expected_source_sha256 and source_sha256 and source_sha256 != expected_source_sha256:
raise ValueError("input CSV checksum differs from the expected snapshot")
source = {
"path": str(input_path),
"bytes": input_path.stat().st_size,
"sha256": source_sha256,
}
schema = _parquet_schema(source, policy, audited_categories)
output_path.parent.mkdir(parents=True, exist_ok=True)
descriptor, temporary_name = tempfile.mkstemp(
dir=output_path.parent, prefix=f".{output_path.name}.", suffix=".tmp"
)
os.close(descriptor)
temporary_path = Path(temporary_name)
total_rows = 0
audited_rows = 0
eligible_rows = 0
malformed_rows = 0
by_category: Counter[str] = Counter()
eligible_by_category: Counter[str] = Counter()
eligibility_rejections: dict[str, Counter[str]] = {}
batch = _new_batch()
try:
with input_path.open("r", encoding="utf-8", newline="") as source_file:
reader = csv.DictReader(source_file)
missing_columns = sorted(REQUIRED_COLUMNS.difference(reader.fieldnames or []))
if missing_columns:
raise ValueError(f"missing required columns: {', '.join(missing_columns)}")
with pq.ParquetWriter(
temporary_path, schema, compression=compression, use_dictionary=True
) as writer:
for row in reader:
total_rows += 1
category = row.get("category") or ""
if category not in category_set:
continue
audited_rows += 1
by_category[category] += 1
try:
eligible = _append_annotation(batch, row, policy)
except (KeyError, TypeError, ValueError):
malformed_rows += 1
continue
if eligible:
eligible_rows += 1
eligible_by_category[category] += 1
else:
rejection_counter = eligibility_rejections.setdefault(
category, Counter()
)
for flag in batch["rejection_flags"][-1]:
rejection_counter[flag] += 1
if len(batch["video_id"]) >= batch_size:
writer.write_table(pa.Table.from_pydict(batch, schema=schema))
batch = _new_batch()
if progress_every_rows and total_rows % progress_every_rows == 0:
if progress_callback:
progress_callback(
total_rows, audited_rows, time.monotonic() - started
)
if batch["video_id"]:
writer.write_table(pa.Table.from_pydict(batch, schema=schema))
os.replace(temporary_path, output_path)
except BaseException:
temporary_path.unlink(missing_ok=True)
raise
report = {
"schema_version": SMART_SEGMENT_ANNOTATION_SCHEMA_VERSION,
"generated_at": datetime.now(tz=UTC).isoformat(),
"normalization_policy": NORMALIZATION_POLICY_VERSION,
"deduplication_policy": deduplication_policy,
"source": source,
"eligibility_policy": asdict(policy),
"categories": audited_categories,
"counts": {
"source_rows": total_rows,
"audited_rows": audited_rows,
"written_rows": audited_rows - malformed_rows,
"eligible_rows": eligible_rows,
"ineligible_rows": audited_rows - malformed_rows - eligible_rows,
"malformed_rows": malformed_rows,
"by_category": dict(sorted(by_category.items())),
"eligible_by_category": dict(sorted(eligible_by_category.items())),
},
"rejections_by_category": {
category: dict(sorted(counter.items()))
for category, counter in sorted(eligibility_rejections.items())
},
"output": {
"path": str(output_path),
"bytes": output_path.stat().st_size,
"sha256": sha256_file(output_path) if compute_sha256 else None,
"compression": compression,
},
"build_seconds": round(time.monotonic() - started, 3),
}
write_json_atomic(manifest_path, report)
return report
@dataclass(frozen=True, slots=True)
class CanonicalAnnotation:
video_id: str
segment_id: str
category: str
start_ms: int
end_ms: int
def load_canonical_annotations(
path: Path,
*,
video_ids: set[str] | None = None,
categories: Sequence[str] | None = None,
eligible_only: bool = True,
batch_size: int = 131_072,
) -> dict[str, list[CanonicalAnnotation]]:
"""Read the canonical annotation table, grouped and sorted by video.
This is the single join point every downstream builder uses instead of
re-interpreting the raw mirror. Cross-category overlaps are returned
intact; each annotation keeps its own category identity.
"""
import pyarrow.parquet as parquet
columns = ["video_id", "segment_id", "category", "start_ms", "end_ms", "is_eligible"]
category_set = set(categories) if categories is not None else None
annotations: dict[str, list[CanonicalAnnotation]] = defaultdict(list)
file = parquet.ParquetFile(path)
available = set(file.schema_arrow.names)
missing = [column for column in columns if column not in available]
if missing:
raise ValueError(
f"canonical annotation table is missing columns: {', '.join(missing)}"
)
for batch in file.iter_batches(columns=columns, batch_size=batch_size):
values = batch.to_pydict()
for video_id, segment_id, category, start_ms, end_ms, is_eligible in zip(
*(values[column] for column in columns), strict=True
):
if video_id is None or start_ms is None or end_ms is None:
continue
if video_ids is not None and video_id not in video_ids:
continue
if category_set is not None and category not in category_set:
continue
if eligible_only and not is_eligible:
continue
annotations[str(video_id)].append(
CanonicalAnnotation(
video_id=str(video_id),
segment_id=str(segment_id) if segment_id is not None else "",
category=str(category),
start_ms=int(start_ms),
end_ms=int(end_ms),
)
)
for records in annotations.values():
records.sort(key=lambda annotation: (annotation.start_ms, annotation.end_ms))
return annotations