shirasko/gemma-2-2b-it-crisp-golf

Unlearned Checkpoint

Field	Value
Unlearning method	CRISP
Base model	google/gemma-2-2b-it
Target concept	Golf
Checkpoint type	LoRA Adapter
Rank / seed	100 / 42
Train eval protocol	mc

---

Unlearning Configuration

Selected hyperparameters (from unlearned_checkpoints.json):

Parameter	Value
alpha	5
delta_embed	0
k_features	20
k_features_embed	0
layer_hi	21
layer_lo	5
layer_step	2
lora_rank	4
lr	0.0005
n_tokens_edited	0
num_epochs	2

---

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

Headline scores used for checkpoint selection:

Metric	Train (after unlearning)	Test (after unlearning)
Efficacy	1	1
Specificity	0.562	0.463
Harmonic mean	0.72	0.633
Relearning QA (MC)	—	0.46

---

Full Evaluation (baseline → unlearned)

From evaluation/score_comparison.csv:

Metric	Baseline (train)	After unlearn (train)	Baseline (test)	After unlearn (test)
QA accuracy	0.86	0.08	0.8	0.02
QA fraction	1	0	1	0
SimDom accuracy	0.82	0.48	0.88	0.44
SimDom fraction	1	0.404	1	0.302
MMLU accuracy	0.52	0.5	0.551	0.554
MMLU fraction	1	0.926	1	1

---

Files in This Repository

File	Description
unlearned_checkpoints.json	Checkpoint metadata & hyperparameters
evaluation/evaluation_summary.json	Full evaluation payload (train/test/relearning)
evaluation/score_comparison.csv	Baseline vs. unlearned comparison table