that has actually been proved to be the case. the weight tweaking being done by abliteration does change token distribution in some cases precisely making LLMs produce tokens it wouldn't have otherwise (and therefore less reflective of training data)