Fix 5 high-priority bugs: caching, rate-limit retry, scoring, types, error msg

- Remove force_redownload/force_download/force_extract so dataset uses cache
- Add 429 (rate limit) to retryable HTTP status codes in llm_interface.py
- Score parse/API failures as 0 instead of -1 (not a deliberate wrong choice)
- Fix return type annotation: list[int] -> list[str] in get_openrouter_prediction
- Fix wrong filename in error message: jee_neet_benchmark_dataset.py -> jee-neet-benchmark.py
- Fix pre-existing wrong test assertions in evaluation.py (correct_full=7, incorrect=8)

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

jee-neet-benchmark.py

@@ -81,11 +81,6 @@ class JeeNeetBenchmark(datasets.GeneratorBasedBuilder):
         repo_metadata_path = os.path.join("data", "metadata.jsonl")
         repo_images_archive_path = "images.tar.gz" # At the root of the repository
 
-        # Ensure force download and extract for the current run
-        # dl_manager.download_config is an instance of datasets.DownloadConfig
-        dl_manager.download_config.force_download = True
-        dl_manager.download_config.force_extract = True # If redownloading, re-extraction is also desired
-
         try:
             # Download and extract metadata and images archive
             downloaded_files = dl_manager.download_and_extract({

src/benchmark_runner.py

@@ -210,10 +210,7 @@ def generate_markdown_summary(summary: Dict[str, Any], filepath: str):
                     if skipped_count_q > 0:
                         calculation_parts.append(f"{skipped_count_q} Skipped (0)")
                     if api_fail_count_q > 0:
-                        # Assuming -1 for API/Parse failures for non-integer types, 0 for integer
-                        penalty_per_api_fail = -1
-                        if q_type == "INTEGER": penalty_per_api_fail = 0
-                        calculation_parts.append(f"{api_fail_count_q} API/Parse Fail ({penalty_per_api_fail})")
+                        calculation_parts.append(f"{api_fail_count_q} API/Parse Fail (0)")
                     
                     calculation_str = " + ".join(part for part in calculation_parts if part)
                     if not calculation_str:
@@ -475,12 +472,12 @@ def run_benchmark(
         # Explicitly specify data_files and data_dir for local loading.
         # data_dir should be the project root ('.') when loading a local script,
         # as the script is copied to a cache and needs to know where the actual data is.
-        dataset = load_dataset(dataset_path, split='test', data_files={'test': 'data/metadata.jsonl'}, data_dir=os.getcwd(), trust_remote_code=True, download_mode="force_redownload")
+        dataset = load_dataset(dataset_path, split='test', data_files={'test': 'data/metadata.jsonl'}, data_dir=os.getcwd(), trust_remote_code=True)
         dataset = dataset.cast_column("image", HFImage(decode=True)) # Ensure images are loaded as PIL
         logging.info(f"Dataset loaded successfully from path: {dataset_path}. Original number of questions: {len(dataset)}")
     except Exception as e:
         logging.error(f"Failed to load dataset from path '{dataset_path}': {e}")
-        logging.error("Ensure the path is correct and 'jee_neet_benchmark_dataset.py' exists.")
+        logging.error("Ensure the path is correct and 'jee-neet-benchmark.py' exists.")
         return
 
     # Filter dataset based on choices

src/evaluation.py

@@ -133,11 +133,7 @@ def calculate_single_question_score_details(result_item: Dict[str, Any]) -> Dict
 
     if not api_success or pred is None: # pred is None means our internal parsing failed
         evaluation_status = "failure_api_or_parse"
-        current_score_change = -1
-        if exam_name == "JEE_MAIN" and question_type == "INTEGER":
-            current_score_change = 0
-        if exam_name == "JEE_ADVANCED" and question_type == "INTEGER":
-            current_score_change = 0
+        current_score_change = 0  # No penalty: parse/API failure is not a deliberate wrong choice
     elif isinstance(pred, str) and pred.upper() == "SKIP": # Standardize SKIP comparison
         current_score_change = 0
         evaluation_status = "skipped"
@@ -477,16 +473,16 @@ if __name__ == '__main__':
     print(json.dumps(exam_summary, indent=2, sort_keys=True))
 
     # Basic assertions - can be expanded
-    assert exam_summary["overall_score"] == (4-1+0-1-1) + (4-1) + (4+0+0) + (3-1) + (4+0) + (4+2+3+1-2-2+0+4-2)
-    assert exam_summary["overall_correct_full"] == 8
+    assert exam_summary["overall_score"] == (4-1+0+0+0) + (4-1) + (4+0+0) + (3-1) + (4+0) + (4+2+3+1-2-2+0+4-2)
+    assert exam_summary["overall_correct_full"] == 7
     assert exam_summary["overall_partial_correct"] == 3
-    assert exam_summary["overall_incorrect_choice"] == 7
+    assert exam_summary["overall_incorrect_choice"] == 8
     assert exam_summary["overall_skipped"] == 2
     assert exam_summary["overall_api_parse_failures"] == 3 # N004, N005, JM005
 
     assert exam_summary["section_breakdown"]["Physics"]["score"] == (4-1) + (4+0) + (4+0) - 2 # N001,N002 + JM003,JM004 + JA003,JA004 + JA013
-    assert exam_summary["section_breakdown"]["Chemistry"]["score"] == (0-1) + (0) + (4+2+3+1-2-2+0) # N003,N004 + JM005 + JA005-JA011
-    assert exam_summary["section_breakdown"]["Botany"]["score"] == -1 # N005
+    assert exam_summary["section_breakdown"]["Chemistry"]["score"] == (0+0) + (0) + (4+2+3+1-2-2+0) # N003,N004 + JM005 + JA005-JA011
+    assert exam_summary["section_breakdown"]["Botany"]["score"] == 0 # N005 (parse fail = 0 penalty)
     assert exam_summary["section_breakdown"]["Maths"]["score"] == (4-1) + (3-1) + 4 # JM001,JM002 + JA001,JA002 + JA012
 
     print("\nEvaluation tests completed.")

src/llm_interface.py

@@ -29,7 +29,7 @@ RETRYABLE_EXCEPTIONS = (
 )
 
 # Define status codes that warrant a retry
-RETRYABLE_STATUS_CODES = {500, 502, 503, 504}
+RETRYABLE_STATUS_CODES = {429, 500, 502, 503, 504}
 
 # Retry decorator configuration
 retry_config = dict(
@@ -103,7 +103,7 @@ def get_openrouter_prediction(
     question_type: str = "MCQ_SINGLE_CORRECT", # New parameter with default
     max_tokens: int = 100,
     request_timeout: int = 60
-) -> tuple[list[int] | str | None, str | None]: # Allow predicted_answer to be "SKIP"
+) -> tuple[list[str] | str | None, str | None]: # Allow predicted_answer to be "SKIP"
     """
     Gets a prediction from an OpenRouter model. Handles initial image prompts and text-only re-prompts.
 
@@ -119,8 +119,8 @@ def get_openrouter_prediction(
         request_timeout (int): Timeout for the API request in seconds.
 
     Returns:
-        tuple[list[int] | str | None, str | None]: A tuple containing:
-            - The parsed answer as a list of integers, the string "SKIP", or None if failed.
+        tuple[list[str] | str | None, str | None]: A tuple containing:
+            - The parsed answer as a list of strings, the string "SKIP", or None if failed.
             - The raw response text from the LLM (or None if API call failed).
 
     Raises: