add politics

LLaVA-MOSS2/add_extra_data.py

@@ -0,0 +1,57 @@
+import json
+import pandas as pd
+
+from llava.train.train import train
+
+with open('./playground/data/llava_v1_5_mix665k.json', 'r', encoding='utf-8') as file:
+    data = json.load(file)
+len = len(data)
+
+paths = ["/root/.cache/huggingface/hub/datasets--dmayhem93--agieval-gaokao-biology/snapshots/91e58112d4022523e02d07cfbc96a950eac9219f/data/test-00000-of-00001-de5aacbbef2a047d.parquet",
+         "/root/.cache/huggingface/hub/datasets--dmayhem93--agieval-gaokao-geography/snapshots/bea8c9da6c3ecf4c07a423b36914f1daa1ba6a1e/data/test-00000-of-00001-82c3eb504d984b0c.parquet",
+         "/root/.cache/huggingface/hub/datasets--dmayhem93--agieval-gaokao-chemistry/snapshots/2fb33cf46ce4aeea9409ea3600a3b1d7e5216536/data/test-00000-of-00001-79e3d766a5e30db5.parquet",
+         "/root/.cache/huggingface/hub/datasets--dmayhem93--agieval-gaokao-chinese/snapshots/d47e4d2c79b7280a7fb9990a11b036dfb8cdd89b/data/test-00000-of-00001-cb21ebb290e0161f.parquet",
+         "/root/.cache/huggingface/hub/datasets--dmayhem93--agieval-gaokao-english/snapshots/691d13566972917f1cdc82f4fa1bad1a5b197cab/data/test-00000-of-00001-8025cecb3b3c0c99.parquet",
+         "/root/.cache/huggingface/hub/datasets--dmayhem93--agieval-gaokao-history/snapshots/41252f835bf3198590df5f4d488d64f78b6fd595/data/test-00000-of-00001-92728bc55381f2f3.parquet",
+         "/root/.cache/huggingface/hub/datasets--dmayhem93--agieval-gaokao-mathqa/snapshots/54ede00f50d90dab8295c0163e16912ee52f8068/data/test-00000-of-00001-31399d80475862e0.parquet",
+         "/root/.cache/huggingface/hub/datasets--dmayhem93--agieval-gaokao-physics/snapshots/3f82847f19ead1a682f0b27cc5c829ac964586bb/data/test-00000-of-00001-d34ffce230cd958a.parquet"]
+datas = []
+for path in paths:
+    df = pd.read_parquet(path)
+
+    dict_list = []
+    for index, row in df.iterrows():
+        dict_item = {}
+        dict_item['id'] = str(len)
+        len+=1
+
+        dict_item['image'] = ""
+
+        conversion = []
+        human = {}
+        human['from'] = 'human'
+        human['value'] = row['query']
+        gpt = {}
+        gpt['from'] = 'gpt'
+        result = "答案是："
+        for option in row['gold']:
+            result += chr(ord('A') + option)
+        gpt['value'] = result
+        conversion.append(human)
+        conversion.append(gpt)
+        dict_item['conversations'] = conversion
+
+        print(dict_item)
+        
+        dict_list.append(dict_item)
+
+    datas = datas + dict_list
+data = data + datas
+with open('data_with_extra_data.json', 'w', encoding='utf-8') as file:
+# 使用json.dump()函数将字典写入文件
+    json.dump(data, file, ensure_ascii=False, indent=4)
+
+new_data = data[::2]
+with open('data_with_extra_data_half.json', 'w', encoding='utf-8') as file:
+# 使用json.dump()函数将字典写入文件
+    json.dump(new_data, file, ensure_ascii=False, indent=4)

LLaVA-MOSS2/llava/serve/submit.py

@@ -1,3 +1,4 @@
+import copy
 import os
 import argparse
 import torch
@@ -5,6 +6,8 @@ import json
 import re
 
 import sys
+
+import tqdm
 sys.path.append('/root/workspace/my-llava-moss2/LLaVA-MOSS2')
 from llava.constants import IMAGE_TOKEN_INDEX, DEFAULT_IMAGE_TOKEN, DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN
 from llava.conversation import conv_templates, SeparatorStyle
@@ -19,6 +22,18 @@ from PIL import Image
 from io import BytesIO
 from transformers import TextStreamer
 
+class SilentStreamer(TextStreamer):
+    def on_finalized_text(self, text: str, stream_end: bool = False):
+        pass
+        # if self.batch_stream:
+        #     if not self.text_cache:
+        #         self.text_cache = text
+        #     elif text:
+        #         self.text_cache = [i + j for i, j in zip(self.text_cache, text)]
+        #     # print(f'\r{self.text_cache}', flush=True, end="" if not stream_end else None)
+        # else:
+        #     # print(text, flush=True, end="" if not stream_end else None)
+        #     pass
 
 def load_image(image_file):
     if image_file.startswith('http://') or image_file.startswith('https://'):
@@ -55,24 +70,6 @@ def get_prompt(key, question, len_of_pictures, image_token):
     D.4
 
     ### 回答：
-    根据欧几里得算法，逐步解析计算两个数6和7的最大公约数（gcd）的步骤如下：
-
-    1. 判断6和7是否相等：不相等。
-    2. 判断6和7大小关系，7 > 6，所以用更大的数7减去较小的数6得到结果1。
-    3. 现在计算6和1的最大公约数。
-    4. 6 > 1，根据算法用更大的数6减去较小的数1得到结果5。
-    5. 再计算5和1的最大公约数。
-    6. 5 > 1，用5减去1得到结果4。
-    7. 再计算4和1的最大公约数。
-    8. 4 > 1，用4减去1得到结果3。
-    9. 再计算3和1的最大公约数。
-    10. 3 > 1，用3减去1得到结果2。
-    11. 再计算2和1的最大公约数。
-    12. 2 > 1，用2减去1得到结果1。
-    13. 最后计算1和1的最大公约数，两数相等，gcd即为这两个数，也就是1。
-
-    因此，6和7的最大公约数是1。
-
     答案是：A.
     
     题目如下：
@@ -136,90 +133,120 @@ def main(args):
     test_data_path = './playground/test'
     questions_path = 'playground/test/questions.json'
     with open(questions_path, 'r', encoding='utf-8') as file:
-        questions = json.load(file)
+        questions_origin = json.load(file)
     
     answer_dic = {'A':0, 'B':0, 'C':0, 'D':0}
 
-    for subject in questions:
-        example = subject['example']
-        for question_itme in example:
-            picture = question_itme['picture']
-            question = question_itme['question']
-            print("question " + str(question_itme['index']) + ":\n")
-    
-            images = [load_image(os.path.join(test_data_path, picture_path)) for picture_path in picture]
-            images_size = [image.size for image in images]
-            image_tensor = process_images(images, image_processor, model.config)
-
-            # image = load_image(args.image_file)
-            # image_size = image.size
-            # # Similar operation in model_worker.py
-            # image_tensor = process_images([image], image_processor, model.config)
-            if type(image_tensor) is list:
-                image_tensor = [image.to(model.device, dtype=torch.float16) for image in image_tensor]
-            else:
-                image_tensor = image_tensor.to(model.device, dtype=torch.float16)
-
-            if len(images) != 0:
-                # first message
-
-                if model.config.mm_use_im_start_end:
-                    # if len(images) == 4:
-                    #     inp = question +'\nA.' + DEFAULT_IM_START_TOKEN + DEFAULT_IMAGE_TOKEN + DEFAULT_IM_END_TOKEN + '\nB.' + DEFAULT_IM_START_TOKEN + DEFAULT_IMAGE_TOKEN + DEFAULT_IM_END_TOKEN + '\nC.' + DEFAULT_IM_START_TOKEN + DEFAULT_IMAGE_TOKEN + DEFAULT_IM_END_TOKEN + '\nD.' + DEFAULT_IM_START_TOKEN + DEFAULT_IMAGE_TOKEN + DEFAULT_IM_END_TOKEN
-                    # elif len(images) == 5:
-                    #     inp = DEFAULT_IM_START_TOKEN + DEFAULT_IMAGE_TOKEN + DEFAULT_IM_END_TOKEN + question +'\nA.' + DEFAULT_IM_START_TOKEN + DEFAULT_IMAGE_TOKEN + DEFAULT_IM_END_TOKEN + '\nB.' + DEFAULT_IM_START_TOKEN + DEFAULT_IMAGE_TOKEN + DEFAULT_IM_END_TOKEN + '\nC.' + DEFAULT_IM_START_TOKEN + DEFAULT_IMAGE_TOKEN + DEFAULT_IM_END_TOKEN + '\nD.' + DEFAULT_IM_START_TOKEN + DEFAULT_IMAGE_TOKEN + DEFAULT_IM_END_TOKEN
-                    # else:
-                    #     inp = DEFAULT_IM_START_TOKEN + DEFAULT_IMAGE_TOKEN + DEFAULT_IM_END_TOKEN + '\n' + question
-                    inp = get_prompt(subject['keyword'], question, len(picture), DEFAULT_IM_START_TOKEN + DEFAULT_IMAGE_TOKEN + DEFAULT_IM_END_TOKEN)
+    answers = []
+
+    for i in tqdm.tqdm(range(0, 5), desc="Voting Processing"):
+        questions = copy.deepcopy(questions_origin)
+        for subject in questions:
+            example = subject['example']
+            for question_itme in tqdm.tqdm(example, desc = f'output_{i}.json ' + subject['keyword'] + ' Processing'):
+                picture = question_itme['picture']
+                question = question_itme['question']
+                # print("question " + str(question_itme['index']) + ":\n")
+        
+                images = [load_image(os.path.join(test_data_path, picture_path)) for picture_path in picture]
+                images_size = [image.size for image in images]
+                image_tensor = process_images(images, image_processor, model.config)
+
+                # image = load_image(args.image_file)
+                # image_size = image.size
+                # # Similar operation in model_worker.py
+                # image_tensor = process_images([image], image_processor, model.config)
+                if type(image_tensor) is list:
+                    image_tensor = [image.to(model.device, dtype=torch.float16) for image in image_tensor]
+                else:
+                    image_tensor = image_tensor.to(model.device, dtype=torch.float16)
+
+                if len(images) != 0:
+                    # first message
+
+                    if model.config.mm_use_im_start_end:
+                        # if len(images) == 4:
+                        #     inp = question +'\nA.' + DEFAULT_IM_START_TOKEN + DEFAULT_IMAGE_TOKEN + DEFAULT_IM_END_TOKEN + '\nB.' + DEFAULT_IM_START_TOKEN + DEFAULT_IMAGE_TOKEN + DEFAULT_IM_END_TOKEN + '\nC.' + DEFAULT_IM_START_TOKEN + DEFAULT_IMAGE_TOKEN + DEFAULT_IM_END_TOKEN + '\nD.' + DEFAULT_IM_START_TOKEN + DEFAULT_IMAGE_TOKEN + DEFAULT_IM_END_TOKEN
+                        # elif len(images) == 5:
+                        #     inp = DEFAULT_IM_START_TOKEN + DEFAULT_IMAGE_TOKEN + DEFAULT_IM_END_TOKEN + question +'\nA.' + DEFAULT_IM_START_TOKEN + DEFAULT_IMAGE_TOKEN + DEFAULT_IM_END_TOKEN + '\nB.' + DEFAULT_IM_START_TOKEN + DEFAULT_IMAGE_TOKEN + DEFAULT_IM_END_TOKEN + '\nC.' + DEFAULT_IM_START_TOKEN + DEFAULT_IMAGE_TOKEN + DEFAULT_IM_END_TOKEN + '\nD.' + DEFAULT_IM_START_TOKEN + DEFAULT_IMAGE_TOKEN + DEFAULT_IM_END_TOKEN
+                        # else:
+                        #     inp = DEFAULT_IM_START_TOKEN + DEFAULT_IMAGE_TOKEN + DEFAULT_IM_END_TOKEN + '\n' + question
+                        inp = get_prompt(subject['keyword'], question, len(picture), DEFAULT_IM_START_TOKEN + DEFAULT_IMAGE_TOKEN + DEFAULT_IM_END_TOKEN)
+                    else:
+                        # if len(images) == 4:
+                        #     inp = question +'\nA.' + DEFAULT_IMAGE_TOKEN + '\nB.' + DEFAULT_IMAGE_TOKEN + '\nC.' + DEFAULT_IMAGE_TOKEN + '\nD.' + DEFAULT_IMAGE_TOKEN
+                        # elif len(images) == 5:
+                        #     inp = DEFAULT_IMAGE_TOKEN + question +'\nA.' + DEFAULT_IMAGE_TOKEN + '\nB.' + DEFAULT_IMAGE_TOKEN + '\nC.' + DEFAULT_IMAGE_TOKEN + '\nD.' + DEFAULT_IMAGE_TOKEN
+                        # else:
+                        #     inp = DEFAULT_IMAGE_TOKEN + '\n' + question
+                        inp = get_prompt(subject['keyword'], question, len(picture), DEFAULT_IMAGE_TOKEN)
+
+                    images = None
+
+                conv = conv_templates[args.conv_mode].copy()
+                if "mpt" in model_name.lower():
+                    roles = ('user', 'assistant')
                 else:
-                    # if len(images) == 4:
-                    #     inp = question +'\nA.' + DEFAULT_IMAGE_TOKEN + '\nB.' + DEFAULT_IMAGE_TOKEN + '\nC.' + DEFAULT_IMAGE_TOKEN + '\nD.' + DEFAULT_IMAGE_TOKEN
-                    # elif len(images) == 5:
-                    #     inp = DEFAULT_IMAGE_TOKEN + question +'\nA.' + DEFAULT_IMAGE_TOKEN + '\nB.' + DEFAULT_IMAGE_TOKEN + '\nC.' + DEFAULT_IMAGE_TOKEN + '\nD.' + DEFAULT_IMAGE_TOKEN
-                    # else:
-                    #     inp = DEFAULT_IMAGE_TOKEN + '\n' + question
-                    inp = get_prompt(subject['keyword'], question, len(picture), DEFAULT_IMAGE_TOKEN)
-
-                images = None
-
-            conv = conv_templates[args.conv_mode].copy()
-            if "mpt" in model_name.lower():
-                roles = ('user', 'assistant')
-            else:
-                roles = conv.roles
-            conv.append_message(conv.roles[0], inp)
-            conv.append_message(conv.roles[1], None)
-            prompt = conv.get_prompt()
-
-            input_ids = tokenizer_image_token(inp, tokenizer, IMAGE_TOKEN_INDEX, return_tensors='pt').unsqueeze(0).to(model.device)
-            stop_str = conv.sep if conv.sep_style != SeparatorStyle.TWO else conv.sep2
-            keywords = [stop_str]
-            streamer = TextStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)
-
-            with torch.inference_mode():
-                output_ids = model.generate(
-                    input_ids,
-                    images=image_tensor,
-                    image_sizes=images_size,
-                    do_sample=True if args.temperature > 0 else False,
-                    temperature=args.temperature,
-                    max_new_tokens=args.max_new_tokens,
-                    streamer=streamer,
-                    use_cache=True)
-
-
-            outputs = tokenizer.decode(output_ids[0]).strip()
-            outputs = re.sub(r'\([^()]*\)', '', outputs)
-            outputs = re.sub(r'<s>|</s>', '', outputs)
-            outputs = extract(outputs, answer_dic)
-            conv.messages[-1][-1] = outputs
-            question_itme['model_answer'] = [outputs]
-            question_itme.pop('picture')
-            question_itme.pop('question')
-
-    with open('output.json', 'w', encoding='utf-8') as file:
+                    roles = conv.roles
+                conv.append_message(conv.roles[0], inp)
+                conv.append_message(conv.roles[1], None)
+                prompt = conv.get_prompt()
+
+                input_ids = tokenizer_image_token(inp, tokenizer, IMAGE_TOKEN_INDEX, return_tensors='pt').unsqueeze(0).to(model.device)
+                stop_str = conv.sep if conv.sep_style != SeparatorStyle.TWO else conv.sep2
+                keywords = [stop_str]
+                streamer = SilentStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)
+
+                with torch.inference_mode():
+                    output_ids = model.generate(
+                        input_ids,
+                        images=image_tensor,
+                        image_sizes=images_size,
+                        do_sample=True if args.temperature > 0 else False,
+                        temperature=args.temperature,
+                        max_new_tokens=args.max_new_tokens,
+                        streamer=streamer,
+                        use_cache=True)
+
+
+                outputs = tokenizer.decode(output_ids[0]).strip()
+                outputs = re.sub(r'\([^()]*\)', '', outputs)
+                outputs = re.sub(r'<s>|</s>', '', outputs)
+                outputs = extract(outputs, answer_dic)
+                conv.messages[-1][-1] = outputs
+                question_itme['model_answer'] = [outputs]
+                question_itme.pop('picture')
+                question_itme.pop('question')
+            
+            # print(subject['keyword'] + "finished")
+
+        answers.append(questions)
+
+    final_ans = answers[0]
+    for ans in answers:
+        for i, sub in enumerate(ans):
+            example = sub['example']
+            for j, item in enumerate(example):
+                item_ans = item['model_answer']
+                index = ord(item_ans[0]) - 65
+                if 'count' not in final_ans:
+                    final_ans[i]['example'][j]['count'] = [0] * 4
+                final_ans[i]['example'][j]['count'][index] += 1
+
+    for sub in final_ans:
+        example = sub['example']
+        for item in example:
+            max = 0
+            for i in range(1, 4):
+                if item['count'][i] > item['count'][max]:
+                    max = i
+            item['model_answer'] = str(chr(max + 65))
+            item.pop('count')
+
+    
+    with open('final_answer.json', 'w', encoding='utf-8') as file:
     # 使用json.dump()函数将字典写入文件
-        json.dump(questions, file, ensure_ascii=False, indent=4)
+        json.dump(final_ans, file, ensure_ascii=False, indent=4)
 
     
     # while True:

LLaVA-MOSS2/read_political.py

@@ -0,0 +1,40 @@
+import jsonlines
+import json
+
+with open('data_with_extra_data.json', 'r', encoding='utf-8') as file:
+    data = json.load(file)
+len = len(data)
+
+path = '/root/.cache/huggingface/hub/datasets--RUCAIBox--gaokao-bench/snapshots/49877cf53b6db9c24d7d285161fc12bba2f85d29/test/2010-2022_Political_Science_MCQs.jsonl'
+with open(path, 'r+', encoding='utf-8') as file:
+    dict_list = []
+    for line in jsonlines.Reader(file):
+        dict_item = {}
+        dict_item['id'] = str(len)
+        len += 1
+
+        dict_item['image'] = ""
+
+        conversion = []
+        human = {}
+        human['from'] = 'human'
+        human['value'] = line['question']
+        gpt = {}
+        gpt['from'] = 'gpt'
+        result = line['analysis']
+        result += "答案是：" + ''.join(line['answer'])
+        gpt['value'] = result
+        conversion.append(human)
+        conversion.append(gpt)
+        dict_item['conversations'] = conversion
+
+        print(dict_item)
+        
+        dict_list.append(dict_item)
+
+    final_data = dict_list + data
+
+with open('political_data_with_extra_data.json', 'w', encoding='utf-8') as file:
+# 使用json.dump()函数将字典写入文件
+    json.dump(data, file, ensure_ascii=False, indent=4)
+

LLaVA-MOSS2/scripts/finetune.sh

@@ -15,7 +15,7 @@ deepspeed llava/train/train_mem.py \
     --deepspeed ./scripts/zero2.json \
     --model_name_or_path /root/.cache/huggingface/hub/models--fnlp--moss2-2_5b-chat/snapshots/3eda5a066c519990bf5f9ba056f5f8ef81531c83 \
     --version $PROMPT_VERSION \
-    --data_path ./data_with_extra_data.json\
+    --data_path ./data_with_extra_data_half.json\
     --image_folder ./playground/data \
     --vision_tower openai/clip-vit-large-patch14 \
     --pretrain_mm_mlp_adapter ./checkpoints/llava-moss2-2_5b-chat-pretrain/mm_projector.bin \
@@ -23,7 +23,7 @@ deepspeed llava/train/train_mem.py \
     --mm_use_im_start_end False \
     --mm_use_im_patch_token False \
     --bf16 True \
-    --max_steps 10000 \
+    --max_steps 40000 \
     --per_device_train_batch_size 2 \
     --per_device_eval_batch_size 2 \
     --gradient_accumulation_steps 2 \

LLaVA-MOSS2/test.py

@@ -1,80 +1,52 @@
-# import re
-# quesiton = "2. (6 分) 研究表明, 氮氧化物和二氧化硫在形成雾霾时与大气中的氨有关如图所示）。下列叙述错误的是（）\nA. 雾和霾的分散剂相同\nB. 雾霾中含有硝酸铵和硫酸铵\nC. $\\mathrm{NH}_3$ 是形成无机颗粒物的催化剂\nD. 雾霾的形成与过度施用氮肥有关"
-# pattern = re.compile(r'\s([A-D]\.\s.*[^\n])')
-#     # 使用findall查找所有匹配的选项
-# options = pattern.findall(quesiton)
-# print(options)
-
-# options = '\n'.join(f"{'ABCDEFG'[i]}. {'Image'}" for i in range(0, 4))
-# print(options)
-
-# def get_prompt(key, question, len_of_pictures, image_token):
-    
-#     pattern = re.compile(r'\s([A-D]\.\s.*[^\n])')
-#     # 使用findall查找所有匹配的选项
-#     options = pattern.findall(question)
-#     if len(options) == 4 or len(options) == 5:
-#         options = '\n'.join(f"{'ABCDEFG'[i]}. {image_token}" for i in range(0, 4))
-#     else:
-#         options = '\n'.join(options)
-#     question = question.split('A.')[0]
-#     if len_of_pictures == 5 or len(options) == 1:
-#             question = image_token + question
-#     prompt = f"""你是一个{key}专家，擅长解决{key}问题。以下是一个{key}的题目，形式为单项选择题。所有的问题都是（close-world assumption）闭世界假设，即未观测事实都为假。请逐步分析问题并在最后一行输出答案，最后一行的格式为"答案是：A"。题目如下：
-    
-#     ### 问题:
-#     {question}
-
-#     ### 选项：
-#     {options}
-#     """
-#     return  prompt
-
-# def extract(input_text):
-#     ans_pattern = re.compile(r"答案是：(.)", re.S)
-
-#     problems = ans_pattern.findall(input_text)
-#     # print(problems)
-#     if(problems == ''):
-#         return 'A'
-#     return problems[0]
+import pandas as pd
+import os
 import json
-import re
 
-def extract(input_text, answer_dic):
-    ans_pattern = re.compile(r"答案是：(.)", re.S)
+with open('./political_data_with_extra_data.json', 'r', encoding='utf-8') as file:
+    data = json.load(file)
+len = len(data)
 
-    problems = ans_pattern.findall(input_text)
-    # print(problems)
-    answer = ''
-    if(len(problems) == 0 or problems == ''):
-        for char in input_text:
-            if char.isupper():
-                answer = "" + char
-    else:
-        answer = problems[0]
+final_folder = 'playground/data/cmmlu'
 
-    if answer == '':
-        answer = 'A'
-        for option in ['B', 'C', 'D']:
-            if answer_dic[option] < answer_dic[answer]:
-                answer = option
-    answer_dic[answer] += 1
+files = os.listdir(final_folder)
 
-    return answer
-
-
-with open("./output.json", 'r', encoding='utf-8') as file:
-    output = json.load(file)
-
-answer_dic = {'A':0, 'B':0, 'C':0, 'D':0}
-
-for subject in output:
-    example = subject['example']
-    for answer in example:
-        answer['model_answer'] = extract(answer['model_answer'], answer_dic)
+selected_files = ['combined_anatomy.csv','combined_ancient_chinese.csv','combined_arts.csv','combined_chinese_civil_service_exam.csv','combined_chinese_foreign_policy.csv',
+                'combined_chinese_history.csv','combined_college_education.csv', 'combined_college_engineering_hydrology.csv', 'combined_college_mathematics.csv', 'combined_college_medicine.csv',
+                'combined_conceptual_physics.csv','combined_electrical_engineering.csv','combined_elementary_mathematics.csv','combined_food_science.csv',
+                'combined_genetics.csv', 'combined_high_school_biology.csv', 'combined_high_school_chemistry.csv','combined_high_school_geography.csv','combined_high_school_mathematics.csv',
+                'combined_high_school_physics.csv','combined_high_school_politics.csv','combined_legal_and_moral_basis.csv','combined_management.csv','combined_marxist_theory.csv',
+                'combined_modern_chinese.csv','combined_philosophy.csv','combined_virology.csv','combined_world_history.csv']
+cmmlu_list = []
+for file_name in selected_files:
+    path = os.path.join(final_folder, file_name)
+    df = pd.read_csv(path)
+    
+    for index, row in df.iterrows():
+        dict_item = {}
+        dict_item['id'] = str(len)
+        len+=1
+
+        dict_item['image'] = ""
+
+        conversion = []
+        human = {}
+        human['from'] = 'human'
+        question = row['Question'] + '\nA.' + row['A'] + '\nB.' + row['B'] + '\nC.' + row['C'] + '\nD' + row['D'] + '\n'
+        human['value'] = question
+        gpt = {}
+        gpt['from'] = 'gpt'
+        result = "答案是：" + row['Answer']
+        gpt['value'] = result
+        conversion.append(human)
+        conversion.append(gpt)
+        dict_item['conversations'] = conversion
+
+        print(dict_item)
+        
+        cmmlu_list.append(dict_item)      
+    
+data = cmmlu_list + data
 
-with open('output.json', 'w', encoding='utf-8') as file:
+with open('cmmlu_political_data_gaokao.json', 'w', encoding='utf-8') as file:
 # 使用json.dump()函数将字典写入文��
-    json.dump(output, file, ensure_ascii=False, indent=4)
-        
+    json.dump(data, file, ensure_ascii=False, indent=4)

LLaVA-MOSS2/vote.py

@@ -0,0 +1,35 @@
+import json
+
+answers = []
+for i in range(1, 5):
+    filename = f'output_{i}.json'
+    with open(filename, 'r', encoding='utf-8') as file:
+        questions = json.load(file)
+        answers.append(questions)
+    
+final_ans = answers[0]
+for ans in answers:
+    for i, sub in enumerate(ans):
+        example = sub['example']
+        for j, item in enumerate(example):
+            item_ans = item['model_answer']
+            index = ord(item_ans[0]) - 65
+            if 'count' not in final_ans:
+                final_ans[i]['example'][j]['count'] = [0] * 4
+            final_ans[i]['example'][j]['count'][index] += 1
+
+for sub in final_ans:
+    example = sub['example']
+    for item in example:
+        max = 0
+        for i in range(1, 4):
+            if item['count'][i] > item['count'][max]:
+                max = i
+        item['model_answer'] = str(chr(max + 65))
+        item.pop('count')
+
+ 
+with open('final_answer.json', 'w', encoding='utf-8') as file:
+# 使用json.dump()函数将字典写入文件
+    json.dump(final_ans, file, ensure_ascii=False, indent=4)
+