Anshuman Suri

So in that case, the Cleverhans implementation is correct, and theirs is not (in the sense that it is not true to the original attack, and they should keep the...

@anderleich I tried fine-tuning on LibriSpeech recently and realized that for WER computation (in the `compute_metric` function) you should normalize text (as is done in the model-card as an example)...

Hey @Framartin, The gradnorm attack is under-construction (should have mentioned it somewhere- my bad!). We started working on it thinking it would be a nice addition, so pasted some placeholder...

Fixed the docstring and closing this issue for now. We might add a token-based gradient attack in a future version, but please feel free to submit a PR in the...

Hey @shanefeng123 . Thanks for your interest in our work! There's unfortunately no documentation or guide on how to do this, but I did push a general direction in the...

@syhsu Version 0.9 ?! Also, cudnn is already installed. P.S : Both keras and theano are up to date.

Hey @kirklandWater1, These n-gram filtered datasets are indeed the only ones available, since they are the ones we use in our work. You can use the `ngram_13_0.8` split if you...

Hey @guangyaodou - this was a deliberate inclusion at the time of code development, since Llama is pretty big and we didn't want scripts to run into OOM issues with...

Hi @JasonH0810 - I used Python 3.9.7 in my setup, with `torch==2.0.1` and `transformers==4.39.1`, but other people that have used the repository have experimented with different (at least later) versions...

Assuming as resolved, closing for now. Re-open if you're still facing issues