CoolFace
Modelpublic

EleutherAI/deep-ignorance-cb-sft-ret2-rm10-orth5

sourceHugging Faceupdated 7mo agoView on Hugging Face
0likes13downloads
Model Card

deep-ignorance-cb-sft-ret2-rm10-orth5

Unlearned model using orthogonal circuit breakers (SFT, full-rank).

Base model: EleutherAI/deep-ignorance-unfiltered

Training hyperparameters

ParameterValue
AlgorithmOrth Circuit Breakers
ModeSFT (full-rank)
retain_coef2
remove_coef10
orth_coef5
Learning rate1e-4

Evaluation results

BenchmarkAccuracy
WMDP Bio Robust0.2903
MMLU0.4440