Add missing paper_id field to dataset features

The JSONL contains paper_id on JEE Advanced rows (198/578). HuggingFace
was failing with CastError because the column existed in data but wasn't
declared in dataset_info features.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

README.md

@@ -36,6 +36,8 @@ dataset_info:
       dtype: string
     - name: correct_answer
       dtype: string
+    - name: paper_id
+      dtype: int64
   splits:
     - name: test
       num_examples: 578

src/benchmark_runner.py

@@ -484,6 +484,7 @@ def run_benchmark(
             for line in f:
                 row = json.loads(line)
                 row["image"] = os.path.join(images_base_dir, row["image"])
+                row.setdefault("paper_id", None)
                 records.append(row)
 
         features = Features({
@@ -494,6 +495,7 @@ def run_benchmark(
             "subject": Value("string"),
             "question_type": Value("string"),
             "correct_answer": Value("string"),
+            "paper_id": Value("int64"),
         })
         dataset = Dataset.from_list(records, features=features)
         logging.info(f"Dataset loaded successfully from {metadata_path}. Original number of questions: {len(dataset)}")