Use concept-erasure implementation of LEACE and SAL #252
Conversation
| @@ -265,12 +265,12 @@ def fit(self, hiddens: Tensor) -> float: | |||
| self.norm.update( | |||
| x=x_neg, | |||
| # Independent indicator for each (template, pseudo-label) pair | |||
| y=torch.cat([torch.zeros_like(prompt_ids), prompt_ids], dim=-1), | |||
| z=torch.cat([torch.zeros_like(prompt_ids), prompt_ids], dim=-1), | |||
There was a problem hiding this comment.
fixed; replaced y with z for ccs
elk/training/eigen_reporter.py
Outdated
There was a problem hiding this comment.
nice :) no need for pseudolabels at inference time
| @@ -40,8 +39,7 @@ def apply_to_layer( | |||
| experiment_dir = elk_reporter_dir() / self.source | |||
|
|
|||
| reporter_path = experiment_dir / "reporters" / f"layer_{layer}.pt" | |||
| reporter = Reporter.load(reporter_path, map_location=device) | |||
| reporter.eval() | |||
There was a problem hiding this comment.
Isn't the eval() here still neded?
There was a problem hiding this comment.
No because CcsReporter doesn't actually have any submodules like nn.BatchNorm or nn.Dropout whose behavior changes due to eval()
|
JFI, my probes / reporters now won't load with this PR because I used Reporter.load. https://github.com/EleutherAI/elk/pull/252/files#diff-d08b84a509f043deeb98c9c642f692fffbd1967486738d2ff242b7897eb0b1ae |
Sorry about that, we can't really guarantee backward compatibility at this point. You should be able to load the reporters with an older commit and extract the raw weights if necessary. |
Now that concept-erasure is on PyPI, we can outsource our
ConceptEraserimplementation to that repo.This PR makes LEACE, rather than SAL, the default method for pseudolabel and prompt template normalization. I should probably add a config option to change it though.