Merge pull request EvolvingLMMs-Lab#37 from EvolvingLMMs-Lab/joshua/o…

…lympiadbench

Joshua/olympiadbench

llava_repr_requirements.txt

@@ -27,6 +27,23 @@ shortuuid==1.0.12
 sqlitedict==2.1.0
 tenacity==8.2.3
 torch==2.0.1
+openai>=1.0.0
+pycocoevalcap
 tokenizers==0.15.2
 tqdm==4.66.2
-transformers==4.37.2
+tqdm-multiprocess
+transformers==4.37.2
+zstandard
+pillow
+pyyaml
+sympy
+mpmath
+Jinja2
+openpyxl
+Levenshtein
+hf_transfer
+tenacity
+wandb>=0.16.0
+transformers-stream-generator
+tiktoken
+pre-commit

lmms_eval/tasks/olympiadbench/cn_utils.py

@@ -0,0 +1,69 @@
+import os
+import json
+import datetime
+from lmms_eval.tasks.olympiadbench.olympiadbench_evals import OlympiadBenchEvaluator
+from lmms_eval.tasks._task_utils.file_utils import generate_submission_file
+
+import logging
+eval_logger = logging.getLogger("lmms-eval")
+dir_name = os.path.dirname(os.path.abspath(__file__))
+
+olympiadbench_evaluator = OlympiadBenchEvaluator()
+
+def olympiadbench_doc_to_visual(doc):
+    return [image.convert("RGB") for image in doc["images"]]
+
+def olympiadbench_doc_to_text(doc):
+    question = doc["question"]
+    subject = doc["subfield"]
+    mul_ans = doc["is_multiple_answer"]
+    if mul_ans is None:
+        mul_ans = False
+    ans_type = doc["answer_type"]
+    if ans_type == "Need_human_evaluate":
+        ans_type = "proof based"
+
+    pre_prompt = f"以下是中国{subject}竞赛中的解答题。\n"
+
+    post_prompt = ""
+    if not mul_ans:
+        post_prompt += f"答案类型为{ans_type}。\n"
+    else:
+        post_prompt += f"题目有多个答案，答案类型均为{ans_type}。\n"
+    post_prompt += "请根据题目的要求和所提供的信息计算得出答案。解答过程和结果中使用的变量和公式请使用LaTeX格式表示。请在最后以"
+    if not mul_ans:
+        post_prompt += '"所以最终答案是\\boxed{答案}。"\n'
+    else:
+        post_prompt += '"所以最终答案是\\boxed{用英⽂逗号连接的多个答案}。"\n'
+
+    final_question = pre_prompt + question + '\n' + post_prompt
+    return final_question
+
+def olympiadbench_process_results(doc, results):
+    precision = doc["error"]
+    is_proving = "TP" in doc["source"] 
+    if precision is None:
+        precision = 0
+    prediction = results[0].strip()
+
+    if is_proving:
+        return {
+            "submission": prediction
+        }
+    else:
+        prediction = prediction.split("所以最终答案是")[-1]
+        prediction = prediction.replace('"', "").replace("\n", "").replace(" ", "").strip(".").strip("。")
+        accuracy = olympiadbench_evaluator.judge(prediction, doc["final_answer"][0], precision)
+        accuracy = int(accuracy)
+        return {
+            "exact_match": accuracy
+        }
+
+def olympiadbench_aggregate_results(results, args):
+    now_date_time = datetime.datetime.now().strftime("%Y-%m%d-%H%M-%S")
+    submission_file_name = f"olympiadbench-test-cn-submission-{now_date_time}.json"
+    path = generate_submission_file(submission_file_name, args)
+    with open(path, "w") as f:
+        json.dump(results, f, ensure_ascii=False)
+    print(f"Submission file saved to {path}")
+

lmms_eval/tasks/olympiadbench/en_utils.py

@@ -0,0 +1,69 @@
+import os
+import json
+import datetime
+from lmms_eval.tasks.olympiadbench.olympiadbench_evals import OlympiadBenchEvaluator
+from lmms_eval.tasks._task_utils.file_utils import generate_submission_file
+
+import logging
+eval_logger = logging.getLogger("lmms-eval")
+dir_name = os.path.dirname(os.path.abspath(__file__))
+
+olympiadbench_evaluator = OlympiadBenchEvaluator()
+
+def olympiadbench_doc_to_visual(doc):
+    return [image.convert("RGB") for image in doc["images"]]
+
+def olympiadbench_doc_to_text(doc):
+    question = doc["question"]
+    subject = doc["subfield"]
+    mul_ans = doc["is_multiple_answer"]
+    if mul_ans is None:
+        mul_ans = False
+    ans_type = doc["answer_type"]
+    if ans_type == "Need_human_evaluate":
+        ans_type = "proof based"
+
+    pre_prompt = f"The following is a question from an International {subject} competition.\n"
+
+    post_prompt = ""
+    if not mul_ans:
+        post_prompt += f"The answer of the question should be {ans_type}.\n"
+    else:
+        post_prompt += f"The question has multiple answers, each of them should be {ans_type}.\n"
+    post_prompt += "Please calculate the answer according to the given requirements and the information provided. Please use LaTeX format to represent the variables and formulas used in the solution process and results. Please end your solution with "
+    if not mul_ans:
+        post_prompt += '"So the final answer is \\boxed{answer}."\n'
+    else:
+        post_prompt += 'So the final answer is \\boxed{multiple answers connected with commas}.\n'
+
+    final_question = pre_prompt + question + '\n' + post_prompt
+    return final_question
+
+def olympiadbench_process_results(doc, results):
+    precision = doc["error"]
+    is_proving = "TP" in doc["source"] 
+    if precision is None:
+        precision = 0
+    prediction = results[0].strip()
+
+    if is_proving:
+        return {
+            "submission": prediction
+        }
+    else:
+        prediction = prediction.split("final answer is")[-1]
+        prediction = prediction.replace('"', "").replace("\n", "").replace(" ", "").strip(".").strip("。")
+        accuracy = olympiadbench_evaluator.judge(prediction, doc["final_answer"][0], precision)
+        accuracy = int(accuracy)
+        return {
+            "exact_match": accuracy
+        }
+
+def olympiadbench_aggregate_results(results, args):
+    now_date_time = datetime.datetime.now().strftime("%Y-%m%d-%H%M-%S")
+    submission_file_name = f"olympiadbench-test-en-submission-{now_date_time}.json"
+    path = generate_submission_file(submission_file_name, args)
+    with open(path, "w") as f:
+        json.dump(results, f, ensure_ascii=False)
+    print(f"Submission file saved to {path}")
+

lmms_eval/tasks/olympiadbench/olympiadbench.yaml

@@ -0,0 +1,6 @@
+group: olympiadbench
+task:
+- olympiadbench_test_en
+- olympiadbench_test_cn
+metadata:
+  - version: 0.0