How to use the sacrebleu.raw_corpus_bleu function in sacrebleu

def test_offset(hypothesis, reference, expected_with_offset, expected_without_offset):
    score_without_offset = sacrebleu.raw_corpus_bleu(hypothesis, reference, 0.0).score / 100
    assert abs(expected_without_offset - score_without_offset) < EPSILON

    score_with_offset = sacrebleu.raw_corpus_bleu(hypothesis, reference, 0.1).score / 100
    assert abs(expected_with_offset - score_with_offset) < EPSILON

def test_offset(hypothesis, reference, expected_with_offset, expected_without_offset):
    score_without_offset = sacrebleu.raw_corpus_bleu(hypothesis, reference, 0.0).score / 100
    assert abs(expected_without_offset - score_without_offset) < EPSILON

    score_with_offset = sacrebleu.raw_corpus_bleu(hypothesis, reference, 0.1).score / 100
    assert abs(expected_with_offset - score_with_offset) < EPSILON

def test_bleu(hypotheses, references, expected_bleu):
    bleu = sacrebleu.raw_corpus_bleu(hypotheses, [references], .01).score / 100
    assert abs(bleu - expected_bleu) < EPSILON

def test_degenerate_uneven(hypotheses, references):
    with pytest.raises(EOFError, match=r'.*stream.*'):
        sacrebleu.raw_corpus_bleu(hypotheses, references)

def test_offset(hypothesis, reference, expected_with_offset, expected_without_offset):
    score_without_offset = sacrebleu.raw_corpus_bleu(hypothesis, reference, 0.0).score / 100
    assert abs(expected_without_offset - score_without_offset) < EPSILON

    score_with_offset = sacrebleu.raw_corpus_bleu(hypothesis, reference, 0.1).score / 100
    assert abs(expected_with_offset - score_with_offset) < EPSILON

def test_offset(hypothesis, reference, expected_with_offset, expected_without_offset):
    score_without_offset = sacrebleu.raw_corpus_bleu(hypothesis, reference, 0.0).score / 100
    assert abs(expected_without_offset - score_without_offset) < EPSILON

    score_with_offset = sacrebleu.raw_corpus_bleu(hypothesis, reference, 0.1).score / 100
    assert abs(expected_with_offset - score_with_offset) < EPSILON

def test_bleu(hypotheses, references, expected_bleu):
    bleu = sacrebleu.raw_corpus_bleu(hypotheses, [references], .01).score / 100
    assert abs(bleu - expected_bleu) < EPSILON

def bleu(hypotheses, references):
    """
    Raw corpus BLEU from sacrebleu (without tokenization)

    :param hypotheses: list of hypotheses (strings)
    :param references: list of references (strings)
    :return:
    """
    return sacrebleu.raw_corpus_bleu(sys_stream=hypotheses,
                                     ref_streams=[references]).score

def raw_corpus_bleu(hypotheses: Iterable[str], references: Iterable[str], offset: Optional[float] = 0.01) -> float:
    """
    Simple wrapper around sacreBLEU's BLEU without tokenization and smoothing.

    :param hypotheses: Hypotheses stream.
    :param references: Reference stream.
    :param offset: Smoothing constant.
    :return: BLEU score as float between 0 and 1.
    """
    return sacrebleu.raw_corpus_bleu(hypotheses, [references], smooth_value=offset).score / 100.0

all_reference_sentences = []
        while not reached_eof:
            batch_x, batch_y, batch_z, batch_tt, actual_batch_size, reached_eof = read_batch_from_samples(dev_samples, self.batch_size, self.token_per_batch, current_index, model.config.data_config.input_features, model.config.data_config.output_features, model.config.data_config.output_translations, model.config.data_config.output_translation_features, model.config.data_config.input_clear_text, model.config.data_config.output_translation_clear_text)
            if actual_batch_size == 0:
                break
            reference = unpad_turn_to_text_and_remove_bpe_of_batch_t(batch_tt[0][0], model.config.data_config.output_translation_vocabularies[0][0])
            for sentence in reference:
                all_reference_sentences.append(sentence)
            output = model.predict_translation_on_batch(batch_x)
            output = unpad_turn_to_text_and_remove_bpe_of_batch_t(output, model.config.data_config.output_translation_vocabularies[0][0])
            for sentence in output:
                all_hypothesis_sentences.append(sentence)
            current_index += actual_batch_size
            if reached_eof is True:
                break
        bleu = sacrebleu.raw_corpus_bleu(sys_stream=all_hypothesis_sentences, ref_streams=[all_reference_sentences])
        return bleu.score