When the phenomenon was first noticed (in GPT2 IIRC), the leading hypothesis was disjunct training sets between the tokenizer and language model—i.e., the anomalous tokens don’t appear in the training set or are poorly represented. It would be strange if this were still the case.
When the phenomenon was first noticed (in GPT2 IIRC), the leading hypothesis was disjunct training sets between the tokenizer and language model—i.e., the anomalous tokens don’t appear in the training set or are poorly represented. It would be strange if this were still the case.