CoolFace
Modelpublic

shirasko/gemma-2-2b-it-snmf-gambling

sourceHugging Faceupdated 2mo agoView on Hugging Face
0likes16downloads
Model Card

Unlearned Checkpoint

FieldValue
Unlearning methodSNMF
Base modelgoogle/gemma-2-2b-it
Target conceptGambling
Checkpoint typeFull Model Weights
Rank / seed100 / 42
Train eval protocolmc

Unlearning Configuration

Selected hyperparameters (from unlearned_checkpoints.json):

ParameterValue
coverage_thresh0.95
delta_embed0
delta_in10
delta_out7
feature_sourceall
k_features_embed0
k_features_mlp_in26
k_features_mlp_out6
layer_hi_in12
layer_hi_out17
layer_lo_in0
layer_lo_out9
n_tokens_edited0
ratio_thresh2
w_modeboth

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

Headline scores used for checkpoint selection:

MetricTrain (after unlearning)**Test (after unlearning)**
Efficacy0.9020.632
Specificity0.8080.738
Harmonic mean0.8520.681
Relearning QA (MC)—0.58

Full Evaluation (baseline → unlearned)

From evaluation/score_comparison.csv:

MetricBaseline (train)After unlearn (train)Baseline (test)**After unlearn (test)**
QA accuracy0.760.30.820.46
QA fraction10.09810.368
SimDom accuracy0.940.780.960.84
SimDom fraction10.76810.831
MMLU accuracy0.520.480.5510.45
MMLU fraction10.85210.664

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