timoschick / self-debiasing Goto Github PK

Hi,

Thanks for open-sourcing the code!

I found that line 220-239 in modeling.py a little bit confusing. Specifically, I have the following questions:

1. Why do we need to flip the attention mask for input prefixes (input_prefixes['attention_mask'] = torch.flip(input_prefixes['attention_mask'], dims=[1]))?
2. Why do we need to roll the input_prefixes['input_ids'] by the length of input_prefixes?
3. From my understanding, we can simply concat the input_prefixes without padding to input_ids_repeated (same for attention mask) and it is done.
4. Why do we need to use shifts[0]? Isn't shifts[0] always 0 because the first prefix is ['']?

Thanks in advance!

apply_dacy_mask convert logits to probability via softmax. In generation.py, the probability is then converted back to logits via torch.log, which may cause numerical instability. In my case, during the ppl evaluation, I encountered some probabilities became 0 and the logits became -inf, which makes the ppl extremely large.

To solve the issue, I wrote an equivalent version below:

def apply_decay_mask_logits(args, logits: torch.Tensor, decay_mask: torch.Tensor) -> torch.Tensor:
    """Applies exponential decay to a tensor of logits"""
    decay_mask = torch.exp(- decay_mask * args.decay_constant)
    decay_mask = torch.max(decay_mask, torch.tensor([args.epsilon], device=decay_mask.device))
    log_decay_mask = torch.log(decay_mask)
    logits += log_decay_mask
    return logits

Please advise. If it looks good to you, I can submit a pull request :)

I found that there are many files in Real Toxicity Prompts Dataset and I would like to know which file was used in your experiments, thanks!

Hi, I noticed that the generate_self_debiasing function is not implemented for the T5 model:

However, in Figure 1 of your paper you give examples of using T5 with self-debiasing.

Would you mind publishing the code for self-debiasing with T5?

Given that T5 is an encoder-decoder model, I assume that self-debiasing has to be performed differently to GPT2, i.e. instead of debiasing the continuation of a prompt, T5 debiases the input sentence itself, or more precisely, the text that is generated for the span in the input sentence that is replaced by a sentinel token. Is it also possible to use self-debiasing with T5 if there are more than one sentinel tokens in the input sentence? Moreover, I'm wondering if it is possible to debias an input sentence with T5 without having to first replace the biased words by sentinel tokens.