CoolFace
Modelpublic

shirasko/gemma-2-2b-it-crisp-wmdp-bio

sourceHugging Faceupdated 2mo agoView on Hugging Face
0likes6downloads
Model Card

Unlearned Checkpoint

FieldValue
Unlearning methodCRISP
Base modelgoogle/gemma-2-2b-it
Target conceptwmdp-bio
Checkpoint typeLoRA Adapter
Rank / seed100 / 42
Train eval protocolmc

Unlearning Configuration

Selected hyperparameters (from unlearned_checkpoints.json):

ParameterValue
alpha5
delta_embed0
k_features5
k_features_embed0
layer_hi15
layer_lo5
layer_step2
lora_rank4
lr0.0005
n_tokens_edited0
num_epochs2

Primary Unlearning Metrics (held-out test, MC protocol)

Headline scores used for checkpoint selection:

MetricTrain (after unlearning)**Test (after unlearning)**
Efficacy10.976
Specificity10.898
Harmonic mean10.935
Relearning QA (MC)—0.68

Full Evaluation (baseline → unlearned)

From evaluation/score_comparison.csv:

MetricBaseline (train)After unlearn (train)Baseline (test)**After unlearn (test)**
QA accuracy0.740.240.660.26
QA fraction1010.024
SimDom accuracy0.620.680.760.7
SimDom fraction1110.882
MMLU accuracy0.520.520.5510.525
MMLU fraction1110.914

Files in This Repository

FileDescription
unlearned_checkpoints.jsonCheckpoint metadata & hyperparameters
evaluation/evaluation_summary.jsonFull evaluation payload (train/test/relearning)
evaluation/score_comparison.csvBaseline vs. unlearned comparison table