kilt_tasks · ELI5 provenances missing in validation split

Jan 19

The provenances present in ELI5 in the original file from the github https://dl.fbaipublicfiles.com/KILT/eli5-dev-kilt.jsonl are missing in the hf_hub dataset.

dmrau changed discussion title from ELI5 provenances missing to ELI5 provenances missing in validation split Jan 19

dmrau

Jan 23

The provenances can be added to the validation split with the following code:

def add_missing_data_eli5(dataset):
    # hf dataset is missing provenances add them
    original_fname = '../data/eli5-dev-kilt.jsonl'
    if not os.path.exists(original_fname):
        urllib.request.urlretrieve("https://dl.fbaipublicfiles.com/KILT/eli5-dev-kilt.jsonl", original_fname)

    original = defaultdict(dict)
    for l in open(original_fname):
        example = json.loads(l)
        original_provenance = [out['provenance'] for out in example['output'] if 'provenance' in out]
        q_id = example['id']
        original[q_id] = original_provenance

    def add_missing_data(x, original_dataset):
        id_ = x['id']
        original_provenance = original_dataset[id_]
        for i in range(len(original_provenance)):
            x['output'][i]['provenance'] = original_provenance[i]
        return x
        
    return dataset.map(add_missing_data, fn_kwargs=dict(original_dataset=original))

dataset = datasets.load_dataset('kilt_tasks', 'eli5' )['validation']
dataset = add_missing_data_eli5(dataset)