Add grouped smoke manifest sampler
Browse files
benchmarks/critic/aesthetic_rm/README.md
CHANGED
|
@@ -31,6 +31,18 @@ python3 reference/benchmarks/critic/aesthetic_rm/filter_multi_candidate_groups.p
|
|
| 31 |
Use `--max-candidates-per-group 4` if a few large groups dominate the manifest
|
| 32 |
and you want a cheaper first pass.
|
| 33 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 34 |
## 2. Run Traditional Metrics
|
| 35 |
|
| 36 |
Use the existing traditional metric runner on the generated metrics manifest:
|
|
|
|
| 31 |
Use `--max-candidates-per-group 4` if a few large groups dominate the manifest
|
| 32 |
and you want a cheaper first pass.
|
| 33 |
|
| 34 |
+
For a tiny grouped smoke test:
|
| 35 |
+
|
| 36 |
+
```bash
|
| 37 |
+
python3 reference/benchmarks/critic/aesthetic_rm/sample_smoke_manifest.py \
|
| 38 |
+
--input out/aesthetic_rm/ditto_candidates_multi.jsonl \
|
| 39 |
+
--output out/aesthetic_rm/smoke_candidates_20.jsonl \
|
| 40 |
+
--metrics-manifest out/aesthetic_rm/smoke_candidates_20_for_metrics.jsonl \
|
| 41 |
+
--groups 10 \
|
| 42 |
+
--candidates-per-group 2 \
|
| 43 |
+
--prefer-remote-paths
|
| 44 |
+
```
|
| 45 |
+
|
| 46 |
## 2. Run Traditional Metrics
|
| 47 |
|
| 48 |
Use the existing traditional metric runner on the generated metrics manifest:
|
benchmarks/critic/aesthetic_rm/sample_smoke_manifest.py
ADDED
|
@@ -0,0 +1,88 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/usr/bin/env python3
|
| 2 |
+
"""Sample a tiny grouped manifest for smoke tests."""
|
| 3 |
+
|
| 4 |
+
from __future__ import annotations
|
| 5 |
+
|
| 6 |
+
import argparse
|
| 7 |
+
import json
|
| 8 |
+
import random
|
| 9 |
+
from collections import defaultdict
|
| 10 |
+
from pathlib import Path
|
| 11 |
+
from typing import Any
|
| 12 |
+
|
| 13 |
+
|
| 14 |
+
def read_jsonl(path: Path) -> list[dict[str, Any]]:
|
| 15 |
+
rows: list[dict[str, Any]] = []
|
| 16 |
+
with path.open("r", encoding="utf-8") as handle:
|
| 17 |
+
for line in handle:
|
| 18 |
+
text = line.strip()
|
| 19 |
+
if text:
|
| 20 |
+
rows.append(json.loads(text))
|
| 21 |
+
return rows
|
| 22 |
+
|
| 23 |
+
|
| 24 |
+
def write_jsonl(path: Path, rows: list[dict[str, Any]]) -> None:
|
| 25 |
+
path.parent.mkdir(parents=True, exist_ok=True)
|
| 26 |
+
with path.open("w", encoding="utf-8") as handle:
|
| 27 |
+
for row in rows:
|
| 28 |
+
handle.write(json.dumps(row, ensure_ascii=False) + "\n")
|
| 29 |
+
|
| 30 |
+
|
| 31 |
+
def make_metrics_rows(rows: list[dict[str, Any]], prefer_remote: bool) -> list[dict[str, Any]]:
|
| 32 |
+
result = []
|
| 33 |
+
for row in rows:
|
| 34 |
+
source = row.get("source_video", "")
|
| 35 |
+
edited = row.get("edited_video", "")
|
| 36 |
+
if prefer_remote:
|
| 37 |
+
source = row.get("remote_source_video") or source
|
| 38 |
+
edited = row.get("remote_edited_video") or edited
|
| 39 |
+
result.append(
|
| 40 |
+
{
|
| 41 |
+
"split": row.get("dataset", ""),
|
| 42 |
+
"sample_id": row["candidate_id"],
|
| 43 |
+
"method": row.get("method", ""),
|
| 44 |
+
"instruction": row.get("instruction", ""),
|
| 45 |
+
"source_video": source,
|
| 46 |
+
"edited_video": edited,
|
| 47 |
+
}
|
| 48 |
+
)
|
| 49 |
+
return result
|
| 50 |
+
|
| 51 |
+
|
| 52 |
+
def main() -> None:
|
| 53 |
+
parser = argparse.ArgumentParser()
|
| 54 |
+
parser.add_argument("--input", type=Path, required=True)
|
| 55 |
+
parser.add_argument("--output", type=Path, required=True)
|
| 56 |
+
parser.add_argument("--metrics-manifest", type=Path, required=True)
|
| 57 |
+
parser.add_argument("--groups", type=int, default=10)
|
| 58 |
+
parser.add_argument("--candidates-per-group", type=int, default=2)
|
| 59 |
+
parser.add_argument("--seed", type=int, default=20260509)
|
| 60 |
+
parser.add_argument("--prefer-remote-paths", action="store_true")
|
| 61 |
+
args = parser.parse_args()
|
| 62 |
+
|
| 63 |
+
rows = read_jsonl(args.input)
|
| 64 |
+
groups: dict[str, list[dict[str, Any]]] = defaultdict(list)
|
| 65 |
+
for row in rows:
|
| 66 |
+
groups[str(row.get("group_id", ""))].append(row)
|
| 67 |
+
|
| 68 |
+
eligible = [
|
| 69 |
+
sorted(group, key=lambda item: (str(item.get("method", "")), str(item["candidate_id"])))
|
| 70 |
+
for group in groups.values()
|
| 71 |
+
if len(group) >= args.candidates_per_group
|
| 72 |
+
]
|
| 73 |
+
rng = random.Random(args.seed)
|
| 74 |
+
rng.shuffle(eligible)
|
| 75 |
+
|
| 76 |
+
selected: list[dict[str, Any]] = []
|
| 77 |
+
for group in eligible[: args.groups]:
|
| 78 |
+
selected.extend(group[: args.candidates_per_group])
|
| 79 |
+
|
| 80 |
+
write_jsonl(args.output, selected)
|
| 81 |
+
write_jsonl(args.metrics_manifest, make_metrics_rows(selected, args.prefer_remote_paths))
|
| 82 |
+
print(f"selected {len(selected)} rows from {min(args.groups, len(eligible))} groups")
|
| 83 |
+
print(f"wrote candidates -> {args.output}")
|
| 84 |
+
print(f"wrote metrics manifest -> {args.metrics_manifest}")
|
| 85 |
+
|
| 86 |
+
|
| 87 |
+
if __name__ == "__main__":
|
| 88 |
+
main()
|