Fix evaluate_from_hf.py: un-pivot columnar annotations, load parquet directly

datasets stores a Sequence(struct) column as a dict of parallel lists rather than a list of per-annotation dicts, so row["annotations"] needed un-pivoting before iterating. Also load the split parquet file(s) directly via snapshot_download + Dataset.from_parquet instead of load_dataset(), since some dataset repos declared dataset_info.features lag behind the actual parquet columns (e.g. this repo is missing reading_order/parent_id/rotation), which made load_dataset raise a cast error.

evaluation_kit/evaluate_from_hf.py

@@ -158,11 +158,19 @@ def annotation_boxes_from_hf_row(row: dict[str, Any]) -> list[AnnotationBox]:
     post-processing has to run here or this entry point would score against a
     different ground truth than the local-JSON scripts do.
     """
+    # `datasets` returns a Sequence(feature=struct) column columnar — a dict
+    # of parallel lists, e.g. {"id": [...], "bbox": [...], ...} — rather than
+    # a list of per-annotation dicts, so it has to be un-pivoted first.
+    columnar = row["annotations"]
+    keys = list(columnar.keys())
+    count = len(columnar[keys[0]]) if keys else 0
+    items = [{key: columnar[key][i] for key in keys} for i in range(count)]
+
     # Rule-labeled regions are decorative separator lines, not text — the
     # raw-JSON loader (load_annotation_boxes) drops them the same way.
     boxes = [
         annotation_box_from_hf_item(item)
-        for item in row["annotations"]
+        for item in items
         if item["label"] != "Rule"
     ]
 
@@ -181,13 +189,28 @@ def main() -> None:
     args = parse_args()
 
     try:
-        from datasets import load_dataset
+        from datasets import Dataset
+        from huggingface_hub import snapshot_download
     except ImportError as exc:
         raise SystemExit(
             "This script needs the `datasets` package: pip install -r requirements.txt"
         ) from exc
 
-    dataset = load_dataset(args.dataset)[args.split]
+    # Load the parquet file(s) directly instead of `datasets.load_dataset(args.dataset)`:
+    # some dataset repos' README-declared `dataset_info.features` lag behind columns
+    # actually present in the parquet data (e.g. RLALT/ACoPDoc's `annotations` struct
+    # is missing `reading_order`/`parent_id`/`rotation` in its declared schema), which
+    # makes `load_dataset` raise a cast error while forcing the data into that stale
+    # schema. `Dataset.from_parquet` infers features from the file itself instead.
+    snapshot_dir = Path(
+        snapshot_download(
+            repo_id=args.dataset, repo_type="dataset", allow_patterns=[f"data/{args.split}-*.parquet"]
+        )
+    )
+    parquet_files = sorted(str(p) for p in snapshot_dir.glob(f"data/{args.split}-*.parquet"))
+    if not parquet_files:
+        raise ValueError(f"{args.dataset}: no data/{args.split}-*.parquet files found")
+    dataset = Dataset.from_parquet(parquet_files)
     if args.dataset_split_column:
         dataset = dataset.filter(
             lambda row: row["split"] == args.dataset_split_column