Upload folder using huggingface_hub

lmms_eval/tasks/mintmcq/utils.py

@@ -16,14 +16,12 @@ from lmms_eval.tasks._task_utils.file_utils import generate_submission_file
 
 VIDEO_TYPE = ["short", "medium", "long"]
 CATEGORIES = ["Artistic Performance", "Culture", "Digital Content", "Knowledge", "Life Record", "Others", "Sports Competition"]
-
-replace_prompt = " Please answer yes or no."
+OPTION_MAP = ['A','B','C','D','E','F','G','H']
 
 with open(Path(__file__).parent / "_default_template_yaml", "r") as f:
     raw_data = f.readlines()
     safe_data = []
     for i, line in enumerate(raw_data):
-        # remove function definition since yaml load cannot handle it
         if "!function" not in line:
             safe_data.append(line)
 
@@ -37,7 +35,6 @@ cache_name = config["dataset_kwargs"]["cache_dir"]
 def convert_time_to_frame(time_in_seconds, fps):
     return int(time_in_seconds * fps)
 
-
 def mint_doc_to_visual(doc):
     cache_dir = os.path.join(base_cache_dir, cache_name) 
     video_path = doc["mint_video_id"]
@@ -56,7 +53,7 @@ def mint_doc_to_visual(doc):
 def mint_doc_to_text(doc, lmms_eval_specific_kwargs=None):
     option_prompt = "Select the best answer to the following multiple-choice question based on the video and the subtitles. Respond with only the letter (A, B, C, or D) of the correct option."
     question = doc["question"]
-    option = "\n".join([f"{opt}" for i, opt in enumerate(doc["options"])])
+    option = "\n".join([f"{OPTION_MAP[i]}. {opt}" for i, opt in enumerate(doc["options"])])
     question = question + "\n" + option
     post_prompt = lmms_eval_specific_kwargs["post_prompt"] if "post_prompt" in lmms_eval_specific_kwargs else "The best answer is:"
     full_prompt = option_prompt + "\n" + question + "\n" + post_prompt
@@ -69,7 +66,8 @@ def extract_characters_regex(s):
         "The correct answer is",
         "The answer is",
         "The answer",
-        "The best option is" "The correct option is",
+        "The best option is",
+        "The correct option is",
         "Best answer:" "Best option:",
     ]
     for answer_prefix in answer_prefixes:
@@ -102,13 +100,9 @@ def mint_process_results(doc, results):
     pred_ans = extract_characters_regex(pred)
 
     category = doc["video_category"]
-    # sub_category = doc["sub_category"]
-    # task_category = doc["task_type"]
-    data_dict = {"question_id": doc["id"], "duration": doc["duration"], "category": category, "pred_answer": pred_ans, "answer": doc["answer"]}
-
-    # return {f"mint_percetion_score": data_dict for metric in matrices}
-    return {f"mint_percetion_score": data_dict}
-
+    doc["answer_option"] = OPTION_MAP[doc["options"].index(doc["answer"])]
+    data_dict = {"question_id": doc["id"], "duration": doc["duration"], "category": category, "pred_answer": pred_ans, "answer": doc["answer"], "answer_option": doc["answer_option"]}
+    return {f"mint_perception_score": data_dict}
 
 def mint_aggregate_results(results):
     """
@@ -127,11 +121,10 @@ def mint_aggregate_results(results):
     for result in results:
         video_type = result["duration"]
         category = result["category"]
-        # sub_category = result["sub_category"]
-        # task_category = result["task_category"]
         key = f"{video_type}_{category}"
         category2score[key]["answered"] += 1
-        category2score[key]["correct"] += result["pred_answer"] == result["answer"]
+        # category2score[key]["correct"] += result["pred_answer"] == result["answer"]
+        category2score[key]["correct"] += result["pred_answer"] == result["answer_option"]
 
     for video_type in VIDEO_TYPE:
         total_correct = 0
@@ -151,24 +144,6 @@ def mint_aggregate_results(results):
                 total_answered += v["answered"]
         eval_logger.info(f"Evaluation on Categories: {category}: {100 * total_correct / total_answered if total_answered > 0 else 0 : .1f}%")
 
-    # for sub_cate in SUB_CATEGORIES:
-    #     total_correct = 0
-    #     total_answered = 0
-    #     for k, v in category2score.items():
-    #         if sub_cate in k:
-    #             total_correct += v["correct"]
-    #             total_answered += v["answered"]
-    #     eval_logger.info(f"Evaluation on Video Sub Categories: {sub_cate}: {100 * total_correct / total_answered if total_answered > 0 else 0 : .1f}%")
-
-    # for task_cate in TASK_CATEGORIES:
-    #     total_correct = 0
-    #     total_answered = 0
-    #     for k, v in category2score.items():
-    #         if task_cate in k:
-    #             total_correct += v["correct"]
-    #             total_answered += v["answered"]
-    #     eval_logger.info(f"Evaluation on Task Categories: {task_cate}: {100 * total_correct / total_answered if total_answered > 0 else 0 : .1f}%")
-
     total_correct = 0
     total_answered = 0
     for k, v in category2score.items():