I load AdvisorQA with this style. You can check all the details in the code.

import json, random

Step 2: Load the dataset

before = []
with open('data/Advice_Collection/v3/train.json') as f:
for line in f:
before.append(line)

after = []
for _, item in enumerate(before):
item = json.loads(item)
num = random.randrange(0,len(item['prefix']))
text = str(item['prefix'][num])[:1000] + '\n\nAnswer: ' + str(item['suffix'][num])
after.append({'text': text})
print(len(after))
after[0]

if tokenizer.pad_token is None:
tokenizer.pad_token = tokenizer.eos_token

ds_train = Dataset.from_list(after)
ds_train = ds_train.shuffle()

tokenizer.pad_token = tokenizer.eos_token
tokenizer.padding_side = "right"