How do you use `is_correct` = False examples from Winogrande etc when training T0
Hi,
Thanks for making the P3 dataset so easy to use! I was wondering how examples where is_correct is False are used when training T0?
Here is an example from the Winogrande validation set where is_correct is False
{'idx': <tf.Tensor: shape=(2,), dtype=int32, numpy=array([0, 0], dtype=int32)>,
'inputs': <tf.Tensor: shape=(34,), dtype=int32, numpy=
array([ 8077, 47, 3, 9, 231, 394, 12177, 145, 6538,
78, 3, 834, 373, 530, 8, 1842, 1488, 5,
363, 405, 8, 3, 834, 16, 8, 756, 7142,
2401, 12, 58, 8077, 42, 6538, 58], dtype=int32)>,
'inputs_pretokenized': <tf.Tensor: shape=(), dtype=string, numpy=b'Sarah was a much better surgeon than Maria so _ always got the easier cases.\nWhat does the _ in the above sentence refer to? Sarah or Maria? '>,
'is_correct': <tf.Tensor: shape=(), dtype=bool, numpy=False>,
'targets': <tf.Tensor: shape=(2,), dtype=int32, numpy=array([8077, 1], dtype=int32)>,
'targets_pretokenized': <tf.Tensor: shape=(), dtype=string, numpy=b'Sarah'>,
'weight': <tf.Tensor: shape=(), dtype=float32, numpy=1.0>}
It seems like we'd only want to train and evaluate with correctly labeled examples with simple text-to-text fine-tuning right?
Hi
@MaximumEntropy
,
the example you are pointing to is coming from a XXX_score_eval subset which is only used for evaluation. We are using the sibling subset XXX (i.e. without "score_eval" appended) for training and which does not contain the is_correct field, and only the correct instances.
Victor
Thanks for clarifying! However, I'm noticing just XXX contains both inputs and targets so I'm not sure I understand the purpose of the XXX_score_eval? Also, how are is_correct=False examples used in validation?
Nw! I hear that i can be confusing the first time :)
For evaluation, we are performing the classification tasks as "rank classification". It means that if an instance has two options or labels (i.e. yes/no, true/false, etc.), we compute the logprob under the model of all options conditioned on the input, and then taking as prediction the option with the highest score.
Essentially, we are comparing logprob(target_A|input), logprob(target_B|input), logprob(target_C|input), ...
For true task zero-shot (i.e. the setup we took a great pain to respect in T0), validation or test does not really matter since technically we are not allowed to use extra information to tune hps/prompts/early stopping/etc. So it's used in the same way I described above.
Let me know if something is not clear!