metadata/data.py

import json
import os

home_dir = '/mnt/bn/bes-mllm-shared/zhangyuan.kevin/LLaVA/playground/data/'
save_dir ='/mnt/bn/bes-mllm-shared/zhangyuan.kevin'

data = json.load(open(os.path.join(home_dir, 'llava_v1_5_mix665k.json')))

# 0 - 364099 coco: 25K
# 364100 - 450516 vg: 20K
# 450517 - 522656 gqa: 20K
# 522657 - 602656 ocrvqa: 20K
# 602657 - 624610 textvqa: 15K
# 624611 - 649999 caption: None

sample_data = []
sample_data.extend(data[0:25000])
sample_data.extend(data[364100:384100])
sample_data.extend(data[450517:470517])
sample_data.extend(data[522657:542657])
sample_data.extend(data[602657:617657])

print('length of reward dataset is:',len(sample_data))

filter_sample_data = []
for item in sample_data:
    i = {}
    i['conversations'] = item['conversations'][:2]
    i['image'] = item['image']
    i['id'] = item['id']

    filter_sample_data.append(i)

with open(os.path.join(save_dir, 'llava_v1_5_mix100k_filter.json'), 'w') as f:
    json.dump(filter_sample_data, f, indent=4)