Support held-out metadata val eval samples

benchmarks/edit/build_eval_samples.py

@@ -1,13 +1,19 @@
 #!/usr/bin/env python3
-"""Build val_N eval sample jsonl files from Ditto manifests.
+"""Build val_N eval sample jsonl files.
 
 The output schema matches the saved val20/val100 artifacts:
 {"id", "target_video", "control_video", "prompt"}.
+
+For reproducing the original val20/val100 split, use --metadata-val-csv with
+one of the held-out real_train/metadata.val.csv files. The manifest sampling
+mode is only a fallback for making a generic eval-like set and is not guaranteed
+to be held out from a specific training run.
 """
 
 from __future__ import annotations
 
 import argparse
+import csv
 import json
 import random
 import re
@@ -39,6 +45,28 @@ def read_json(path: Path) -> list[dict[str, Any]]:
     return data
 
 
+def read_metadata_val_csv(path: Path, count: int) -> list[dict[str, str]]:
+    rows: list[dict[str, str]] = []
+    with path.open("r", encoding="utf-8", newline="") as handle:
+        reader = csv.DictReader(handle)
+        missing = {"video", "vace_video", "prompt"} - set(reader.fieldnames or [])
+        if missing:
+            raise ValueError(f"{path} missing required columns: {sorted(missing)}")
+        for row in reader:
+            rows.append(
+                {
+                    "target_video": str(row["video"]),
+                    "control_video": str(row["vace_video"]),
+                    "prompt": str(row["prompt"]),
+                }
+            )
+            if len(rows) >= count:
+                break
+    if len(rows) < count:
+        raise RuntimeError(f"{path} only has {len(rows)} rows, requested {count}")
+    return rows
+
+
 def sanitize_flat_relpath(relpath: str) -> str:
     parts = [re.sub(r"[^0-9A-Za-z._-]+", "_", piece) for piece in relpath.split("/")]
     return "__".join(parts)
@@ -146,6 +174,12 @@ def main() -> None:
     parser = argparse.ArgumentParser()
     parser.add_argument("--repo-root", type=Path, default=Path.cwd())
     parser.add_argument("--manifest", type=Path, action="append", default=[])
+    parser.add_argument(
+        "--metadata-val-csv",
+        type=Path,
+        default=None,
+        help="Held-out real_train/metadata.val.csv. Use this to reproduce val20/val100 logic.",
+    )
     parser.add_argument("--output", type=Path, required=True)
     parser.add_argument("--count", type=int, default=1000)
     parser.add_argument("--seed", type=int, default=20260511)
@@ -153,14 +187,20 @@ def main() -> None:
     parser.add_argument("--bucket-prefix", action="append", default=[])
     args = parser.parse_args()
 
-    manifests = args.manifest or [Path(p) for p in DEFAULT_MANIFESTS]
-    bucket_prefixes = tuple(args.bucket_prefix or DEFAULT_BUCKET_PREFIXES)
-    candidates = collect_candidates(args.repo_root, manifests, bucket_prefixes, args.path_style)
-    selected = stratified_sample(candidates, args.count, args.seed)
+    if args.metadata_val_csv is not None:
+        metadata_path = args.metadata_val_csv if args.metadata_val_csv.is_absolute() else args.repo_root / args.metadata_val_csv
+        selected = read_metadata_val_csv(metadata_path, args.count)
+        print(f"source={metadata_path}")
+        print("mode=metadata.val.csv prefix")
+    else:
+        manifests = args.manifest or [Path(p) for p in DEFAULT_MANIFESTS]
+        bucket_prefixes = tuple(args.bucket_prefix or DEFAULT_BUCKET_PREFIXES)
+        candidates = collect_candidates(args.repo_root, manifests, bucket_prefixes, args.path_style)
+        selected = stratified_sample(candidates, args.count, args.seed)
+        print(f"candidates={len(candidates)}")
+        print(f"path_style={args.path_style}, seed={args.seed}")
     write_samples(args.output, selected)
-    print(f"candidates={len(candidates)}")
     print(f"selected={len(selected)} -> {args.output}")
-    print(f"path_style={args.path_style}, seed={args.seed}")
 
 
 if __name__ == "__main__":