@lukastymo@adamzwasserman I thought the same. The convict for this behavior should be RL, not pretraining. However, some pretrained checkpoint had a similar or higher probability mass on number 7 then its instruction tuned version. This might be due to my prompting though.
@wufeima@diyerxx None of the ICLR reviewers caught that problem, and if not notified, the faulty dataset might have been published
> But why post this in a way that points fingers at the authors?
Agree with this
3. Allow users to block attention from and to certain tokens.
4. Save the current intervention/outputs into JSON and npz files for further inspection.
Github link: https://t.co/O6RAGDA9fz
Hi all. I created an interactive logit lens for Llama models, something I wished existed a year ago. It can:
1. Interactively show logit lens output for user input
2. Allow users to modify the residual stream, attention outputs, and MLP outputs