README.md

Loading the KILT knowledge source and task data

The original KILT release only provides question IDs for the TriviaQA task. Using the full dataset requires mapping those back to the TriviaQA questions, which can be done as follows:

from datasets import load_dataset

# Get the pre-processed Wikipedia knowledge source for kild
kilt_wiki = load_dataset("kilt_wikipedia")

# Get the KILT task datasets
kilt_tasks = load_sataset("kilt_tasks")

# Most tasks in KILT already have all required data, but KILT-TriviaQA
# only provides the question IDs, not the questions themselves.
# Thankfully, we can get the original TriviaQA data with:
trivia_qa = load_dataset('trivia_qa', 'unfiltered.nocontext')

# The KILT IDs can then be mapped to the TriviaQA questions with:
triviaqa_map = {}
for k in ['train', 'validation', 'test']:
    triviaqa_map = dict([(q_id, i) for i, q_id in enumerate(triviaqa[k]['question_id'])])
    kilt_tasks[k + '_triviaqa'] = kilt_tasks[k + '_triviaqa'].filter(lambda x: x['id'] in triviaqa_map)
    kilt_tasks[k + '_triviaqa'] = kilt_tasks[k + '_triviaqa'].map(lambda x: {'input': triviaqa[k][triviaqa_map[x['id']]]['question']})