eleutherai / concept-erasure Goto Github PK

Hi,

Reporting a bug with the latest package version concept-erasure 0.2.1, python 10.4 and torch 1.13.0+cu116.

The second argument to torch.Tensor.where() has to be a tensor, not a float. Fixed it with the following

Lzeros = torch.zeros(L.shape) W = V * L.rsqrt().where(mask, Lzeros) @ V.mT W_inv = V * L.sqrt().where(mask, Lzeros) @ V.mT

This was the only code that executed in the runtime, and the same thing happens for both CUDA and CPU instances.

Am I wrong in thinking QuadraticEraser is supposed to work the same as LeaceEraser?

Adapting the code from:

!pip install concept-erasure datasets transformers --quiet

from itertools import pairwise, product

import numpy as np
import pytest
import torch
import torch.nn.functional as F
from sklearn.datasets import make_classification
from sklearn.discriminant_analysis import QuadraticDiscriminantAnalysis
from sklearn.linear_model import LogisticRegression
from sklearn.metrics import log_loss
from sklearn.svm import LinearSVC

from concept_erasure import (
    ErasureMethod,
    LeaceFitter,
    OracleEraser,
    OracleFitter,
    QuadraticEraser,
    optimal_linear_shrinkage,
)
from concept_erasure.optimal_transport import is_positive_definite

In many use cases you don't need transformers when using concept_erasure, and transformers is a heavyweight library so this adds a lot to the import time. More generally transformers could probably be made an optional dependency.

Hi, thanks for repository and paper. Is it possible to apply this to generation tasks in language models and not just classification ?
I am very interested in this aspect. Also, just to confirm, the scrubber is a technique that is applied during inference and doesn't modify model parameters right ? It only modifies hidden representations ?

Hi, would it be possible to provide a very simply sample to patch a llama model removing a specific singular text concept from the model? The sample provided on the README is slightly confusing