add Min-K%++ attack #19
Conversation
|
Thanks a lot for the PR, @zjysteven ! I have a couple suggestions that I have added via comments. The fact that you submitted such a well-structured PR is alone sufficient and I would understand if you would rather have us make these changes, but would be great if you could make those changes |
|
Yes I can make further changes. Somehow I'm not seeing your comments. Would you please point me to the comments? |
| target_prob, all_prob = ( | ||
| (probs, all_probs) | ||
| if (probs is not None and all_probs is not None) | ||
| else self.model.get_probabilities(document, tokens=tokens, return_all_probs=True) |
There was a problem hiding this comment.
Is there a reason for computing softmax and log_softmax separately? Given access to the latter (which is default right now), a exp() operation should give you the desired softmax outputs for use in mink++?
There was a problem hiding this comment.
I see. Earlier I misread and thought that we have softmax, and since log_softmax is numerically more stable than log(softmax) I compute them separately (but yeah as you pointed what we have already is log_softmax). I will make the change.
There was a problem hiding this comment.
I do want to clarify here that while target_prob is the log probability of each input token, all_prob is the log probabilities of every token in the vocabulary (the whole categorical distribution).
run.py
Outdated
| loss=loss, | ||
| ) | ||
| sample_information[attack].append(score) | ||
| if attack in [AllAttacks.MIN_K, AllAttacks.MIN_K_PLUS_PLUS]: |
There was a problem hiding this comment.
Our intent with the structure was to have as few attack-specific checks as possible. We plan on adding attack-specific hyper-parameters in the config file, so that any changes required for the attack here can be achieved by different config files, and not hard-coded hyper-parameters and switch-case in code!
There was a problem hiding this comment.
Also a valid point. I will just remove this modification so run.py follows the current design.
|
Thanks! They should be visible now |
|
I've made changes accordingly in the latest commit. @iamgroot42 Would you take another look? |
| @@ -116,7 +113,6 @@ def get_probabilities(self, | |||
| if no_grads: | |||
| logits = logits.cpu() | |||
| shift_logits = logits[..., :-1, :].contiguous() | |||
| probabilities = torch.nn.functional.softmax(shift_logits, dim=-1) | |||
There was a problem hiding this comment.
Like suggested here we don't need to separately call softmax given that log_softmax will be computed anyway.
|
Thanks for making the changes! Will run our standard workflow and merge it |
Hi @iamgroot42, first thanks for putting up this unified benchmark.
Overview
This pull request add our proposed method, Min-K%++, which outperforms other reference-free methods and performs on par with the Ref method on MIMIR. See below experiment part for details.
Changed files
We follow the instruction to create
attacks/mink_plus_plus.pyand register the method inattacks/all_attacks.pyandattacks/utils.py. Our method needs the full probabilities over the vocabulary, so we modify theget_probabilitiesfunction inmodels.pywith an additional argumentreturn_all_probsto return full probs.Meanwhile, we change
run.pyaccordingly to include our Min-K%++ attack. Additionally, we let both Min-K% and Min-K%++ to run over a list ofkvalues, which allows both methods to be compared by their upper bounds.Lastly, we change the readme to incorporate Min-K%++.
Experiments
I'm putting this result table here in case MIMIR wants to have a leaderboard or something. We are using non-deduped pythia models and the default
ngram_13_0.8setting. The data splits are fetched fromcache_100_200_1000_512folder on huggingface.Let us know if there are any issues. Otherwise, we look forward to Min-K%++ being integrated into the MIMIR benchmark.