EleutherAI / elk

Keeping language models honest by directly eliciting knowledge encoded in their activations.
MIT License
178 stars 33 forks source link

`burns` shortcut dataset in sweep #236

Closed norabelrose closed 1 year ago

norabelrose commented 1 year ago

Adds a magic burns dataset which expands to all the datasets used in Burns et al. (2022) except Story Cloze, which is not on the HF Hub and may be ~impossible to access at the moment since the form is broken.