CoolFace
Modelpublic

shirasko/qwen3.5-2b-snmf-uranium

sourceHugging Faceupdated 21d agoView on Hugging Face
0likes460downloads
Model Card

Unlearned Checkpoint

FieldValue
Unlearning methodSNMF
Base modelQwen/Qwen3.5-2B
Target conceptUranium
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_in7
delta_out1
feature_sourceall
k_features_embed0
k_features_mlp_in20
k_features_mlp_out9
layer_hi_in12
layer_hi_out16
layer_lo_in0
layer_lo_out8
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)**
Efficacy0.6270.824
Specificity0.5940.444
Harmonic mean0.610.577
Relearning QA (MC)—0.42

Full Evaluation (baseline → unlearned)

From evaluation/score_comparison.csv:

MetricBaseline (train)After unlearn (train)Baseline (test)**After unlearn (test)**
QA accuracy0.760.440.760.34
QA fraction10.37310.176
SimDom accuracy0.70.440.760.4
SimDom fraction10.42210.294
MMLU accuracy0.560.580.5880.556
MMLU fraction1110.905

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