bart-dataset-v3 / scripts /run_filter.py
jbduran's picture
upload/update 1930s pipeline scripts
bcedfe6 verified
Raw
History Blame Contribute Delete
7.47 kB
"""Stage 3: main resumable shard loop.
Downloads (or builds) the banned list, compiles the fast matcher, then for each
source shard: scans every whole-book row, drops docs with anachronism hits, and
writes a same-named output shard + per-shard stats + a per-shard hit log, all in
one commit. Resumes by skipping shards already present in the destination repo.
Prints a per-shard ETA from a running average.
Run: python scripts/run_filter.py
First-shard dry-run: DRY_RUN_LIMIT=1 python scripts/run_filter.py
"""
import json
import time
from collections import defaultdict
from pathlib import Path
import pyarrow as pa
import pyarrow.parquet as pq
from huggingface_hub import CommitOperationAdd
import config
import filter_lib
from build_list import load_or_build
from common import (
BatchCommitter,
destination_shards,
download_source_shard,
ensure_dst_repo,
output_path_for_source,
source_shards,
)
def write_parquet_texts(texts, local_out):
table = pa.table({"text": texts})
pq.write_table(
table, local_out,
row_group_size=config.ROW_GROUP_SIZE,
compression=config.COMPRESSION,
compression_level=config.COMPRESSION_LEVEL,
)
def fmt_secs(s):
m, s = divmod(int(s), 60)
h, m = divmod(m, 60)
return f"{h}h{m:02d}m" if h else f"{m}m{s:02d}s"
def process_shard(path_in_repo):
output_name = output_path_for_source(path_in_repo)
stem = Path(output_name).stem
local_source = download_source_shard(path_in_repo)
local_out = config.OUT_DIR / output_name
local_stats = config.OUT_DIR / f"{stem}.json"
local_hits = config.OUT_DIR / f"{stem}.jsonl"
stats = {
"source_shard": path_in_repo,
"output_shard": output_name,
"started_at": time.strftime("%Y-%m-%dT%H:%M:%SZ", time.gmtime()),
"n_input": 0, "n_kept": 0, "n_removed": 0,
"removed_by_term": defaultdict(int), # terms that caused drops
"removed_by_reason": defaultdict(int), # tier1 vs corroborated_tier2
"boilerplate_hits": defaultdict(int), # strip-only signals seen (kept docs)
"chars_in": 0, "chars_kept": 0,
}
kept_texts, hit_records = [], []
pf = pq.ParquetFile(local_source)
doc_idx = -1
for rg_idx in range(pf.num_row_groups):
table = pf.read_row_group(rg_idx, columns=["text"])
for text in table.column("text").to_pylist():
doc_idx += 1
stats["n_input"] += 1
stats["chars_in"] += len(text) if isinstance(text, str) else 0
terms, kinds = filter_lib.scan_text(text)
drop, reason, by_tier = filter_lib.decide_drop(terms)
if drop:
stats["n_removed"] += 1
stats["removed_by_reason"][reason] += 1
# Only the tiered signals that actually justified the drop.
for t in by_tier[1] + by_tier[2] + by_tier[3]:
stats["removed_by_term"][t] += 1
hit_records.append({
"doc_idx": doc_idx, "reason": reason,
"tier1": by_tier[1], "tier2": by_tier[2],
"tier3": by_tier[3], "strip": by_tier["strip"],
})
continue
# Kept doc: still note any boilerplate/strip-only signals for auditing.
for t in by_tier["strip"]:
stats["boilerplate_hits"][t] += 1
kept_texts.append(text) # unchanged -- whole book preserved
stats["n_kept"] += 1
stats["chars_kept"] += len(text) if isinstance(text, str) else 0
stats["removed_by_term"] = dict(stats["removed_by_term"])
stats["removed_by_reason"] = dict(stats["removed_by_reason"])
stats["boilerplate_hits"] = dict(stats["boilerplate_hits"])
stats["finished_at"] = time.strftime("%Y-%m-%dT%H:%M:%SZ", time.gmtime())
write_parquet_texts(kept_texts, local_out)
with open(local_stats, "w", encoding="utf-8") as f:
json.dump(stats, f, indent=2, sort_keys=True)
with open(local_hits, "w", encoding="utf-8") as f:
for rec in hit_records:
f.write(json.dumps(rec) + "\n")
# Return the upload operations; the caller batches them into one commit.
ops = [
CommitOperationAdd(output_name, str(local_out)),
CommitOperationAdd(f"stats/{stem}.json", str(local_stats)),
CommitOperationAdd(f"hits/{stem}.jsonl", str(local_hits)),
]
return stats, ops
def main():
ensure_dst_repo()
config.ensure_dirs()
banned_terms, meta, tiers = load_or_build()
filter_lib.init_matcher(banned_terms, tiers)
tc = (meta or {}).get("tier_counts", {})
print(f"Matcher ready: {len(banned_terms):,} banned terms "
f"(T1={tc.get('tier1', 0)} T2={tc.get('tier2', 0)} T3={tc.get('tier3', 0)} "
f"strip={tc.get('strip', 0)}), {len(filter_lib.FORMAT_RES)} format tells. "
f"Drop rule: 1x T1 OR >=2 T2/T3 with >=1 T2. "
f"Scan window: head={config.SCAN_CHARS or 'ALL'} tail={config.SCAN_TAIL_CHARS}.")
# Read shards from SRC_REPO/SRC_PREFIX -- point these at the stripped layer:
# SRC_REPO=<dst repo> SRC_PREFIX=stripped
all_source = source_shards(prefix=config.SRC_PREFIX)
done = destination_shards(root_only=True) # final shards live at repo root only
done_names = {Path(p).name for p in done}
todo = [s for s in all_source if output_path_for_source(s) not in done_names]
src_desc = f"{config.SRC_REPO}" + (f"/{config.SRC_PREFIX}/" if config.SRC_PREFIX else "")
print(f"Filter input: {src_desc}")
print(f"Source shards: {len(all_source):,} | already done: {len(all_source) - len(todo):,} | remaining: {len(todo):,}")
if config.DRY_RUN_LIMIT:
todo = todo[:config.DRY_RUN_LIMIT]
print(f"DRY_RUN_LIMIT={config.DRY_RUN_LIMIT}: processing only {len(todo)} shard(s) this run.")
committer = BatchCommitter(
repo_id=config.DST_REPO, batch_size=config.BATCH_SIZE,
message_prefix="filter",
)
print(f"Batching uploads: {config.BATCH_SIZE} shard(s) per commit.")
t0 = time.time()
n_total = len(todo)
for i, shard in enumerate(todo, start=1):
st, ops = process_shard(shard)
committed = committer.add(ops, label=st["output_shard"])
removed_pct = 100.0 * st["n_removed"] / max(st["n_input"], 1)
elapsed = time.time() - t0
avg = elapsed / i
eta = avg * (n_total - i)
top = sorted(st["removed_by_term"].items(), key=lambda x: -x[1])[:5]
flush_note = f" [committed batch #{committer.n_commits}]" if committed else f" [buffered {committer.pending}]"
print(
f"[{i}/{n_total}] {st['output_shard']}: kept {st['n_kept']:,}/{st['n_input']:,} "
f"removed={st['n_removed']:,} ({removed_pct:.2f}%) | "
f"elapsed {fmt_secs(elapsed)} | avg {avg:.1f}s/shard | "
f"ETA ~{fmt_secs(eta)} for {n_total - i} left | top={top}{flush_note}"
)
committer.flush() # push the final partial batch
print(f"\nThis run processed {len(todo):,} shard(s) in {fmt_secs(time.time() - t0)} "
f"({committer.n_commits} commit(s)).")
if config.DRY_RUN_LIMIT and (len(all_source) - len(done)) > len(todo):
print("Dry run complete. Inspect hits/ on HF, then unset DRY_RUN_LIMIT and re-run "
"to process the rest (completed shards are skipped).")
if __name__ == "__main__":
main()