MemeRewardModelDataset / data_to_llamafactory.py
Karroyan's picture
Upload data_to_llamafactory.py with huggingface_hub
a66ea25 verified
# turn the data in /fs-computility/niuyazhe/lixueyan/meme/dataset-meme-rewardmodel/allmeme_nocross/allmeme_nocross_train.jsonl
# to the format in /fs-computility/niuyazhe/lixueyan/meme/LLaMA-Factory/data/binary_class_demo.json
import json
import jsonlines
import os
output_path_root = '/fs-computility/niuyazhe/lixueyan/meme/LLaMA-Factory/data/'
with open('/fs-computility/niuyazhe/lixueyan/meme/dataset-meme-rewardmodel/allmeme_nocross/allmeme_nocross_train.jsonl', 'r') as f:
dataset_dict = json.load(f)
for dataset_name, dataset in dataset_dict.items():
dataset_path = dataset['annotation']
dataset_name = dataset_path.split('/')[-1].split('.')[0]
# dataset_path is also jsonl file
with open(dataset_path, 'r') as file:
for item in jsonlines.Reader(file):
# check if the image exists
if not os.path.exists(item['image'][0]) or not os.path.exists(item['image'][1]):
print(f"image {item['image'][0]} or {item['image'][1]} does not exist, skip this item")
# skip the whole item
continue
new_item = {}
new_item['messages'] = [
{
'role': 'user',
'content': item['conversations'][0]['value']
},
{
'role': 'assistant',
'content': ''
}
]
new_item['images'] = item['image']
new_item['labels'] = [item['conversations'][1]['value']]
with open(output_path_root + dataset_name + '.json', 'a') as f:
json.dump(new_item, f)
f.write('\n')