Sync DeepSeek OCR HF job code

ds_batch_ocr/stages.py

@@ -41,30 +41,43 @@ def write_jsonl(path: Path, rows: List[Dict[str, Any]]) -> None:
 
 
 def _normalize_figures(figures: Any) -> List[Dict[str, str]]:
+    def _unwrap(entry: Any) -> Any:
+        current = entry
+        # peel off single-item wrappers (e.g. [[{...}]]) common in some datasets
+        while isinstance(current, (list, tuple)) and len(current) == 1:
+            nested = current[0]
+            if isinstance(nested, (dict, list, tuple)) or hasattr(nested, "figure_id"):
+                current = nested
+            else:
+                break
+        return current
+
     def _to_dict(entry: Any) -> Dict[str, str]:
         figure_id = ""
         image_path = ""
         description = ""
 
-        if isinstance(entry, dict):
-            figure_id = entry.get("figure_id") or entry.get("id") or ""
+        normalized_entry = _unwrap(entry)
+
+        if isinstance(normalized_entry, dict):
+            figure_id = normalized_entry.get("figure_id") or normalized_entry.get("id") or ""
             image_path = (
-                entry.get("image_path")
-                or entry.get("document_relative_path")
-                or entry.get("path")
+                normalized_entry.get("image_path")
+                or normalized_entry.get("document_relative_path")
+                or normalized_entry.get("path")
                 or ""
             )
-            description = entry.get("description") or entry.get("caption") or ""
-        elif hasattr(entry, "figure_id"):
-            figure_id = getattr(entry, "figure_id", "")
-            image_path = getattr(entry, "image_path", "") or getattr(entry, "document_relative_path", "")
-            description = getattr(entry, "description", "")
-        elif isinstance(entry, (list, tuple)):
-            figure_id = entry[0] if len(entry) > 0 else ""
-            image_path = entry[1] if len(entry) > 1 else ""
-            description = entry[2] if len(entry) > 2 else ""
+            description = normalized_entry.get("description") or normalized_entry.get("caption") or ""
+        elif hasattr(normalized_entry, "figure_id"):
+            figure_id = getattr(normalized_entry, "figure_id", "")
+            image_path = getattr(normalized_entry, "image_path", "") or getattr(normalized_entry, "document_relative_path", "")
+            description = getattr(normalized_entry, "description", "")
+        elif isinstance(normalized_entry, (list, tuple)):
+            figure_id = normalized_entry[0] if len(normalized_entry) > 0 else ""
+            image_path = normalized_entry[1] if len(normalized_entry) > 1 else ""
+            description = normalized_entry[2] if len(normalized_entry) > 2 else ""
         else:
-            figure_id = entry
+            figure_id = normalized_entry
 
         return {
             "figure_id": str(figure_id or ""),
@@ -202,7 +215,25 @@ def _push_dataset_records(
                 }
             )
 
-        new_record["figures"] = coerced_figures
+        clean_figures: List[Dict[str, str]] = []
+        for fig in coerced_figures:
+            if not isinstance(fig, dict):
+                LOGGER.warning(
+                    "Dropping non-dict figure entry after coercion | type=%s | value=%r",
+                    type(fig),
+                    fig,
+                )
+                continue
+
+            clean_figures.append(
+                {
+                    "figure_id": str(fig.get("figure_id", "") or ""),
+                    "image_path": str(fig.get("image_path", "") or ""),
+                    "description": str(fig.get("description", "") or ""),
+                }
+            )
+
+        new_record["figures"] = clean_figures
         normalized_records.append(new_record)
 
     dataset = Dataset.from_list(normalized_records, features=_dataset_features())