shirasko/llama-3.1-8b-instruct-crisp-golf

Unlearned Checkpoint

Field	Value
Unlearning method	CRISP
Base model	meta-llama/Llama-3.1-8B-Instruct
Target concept	Golf
Checkpoint type	LoRA Adapter
Rank / seed	200 / 42
Train eval protocol	mc

---

Unlearning Configuration

Selected hyperparameters (from unlearned_checkpoints.json):

Parameter	Value
alpha	50
delta_embed	0
k_features	20
k_features_embed	0
layer_hi	28
layer_lo	4
layer_step	2
lora_rank	4
lr	0.0001
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	0.881	0.889
Specificity	0.786	0.914
Harmonic mean	0.831	0.901
Relearning QA (MC)	—	0.8

---

Full Evaluation (baseline → unlearned)

From evaluation/score_comparison.csv:

Metric	Baseline (train)	After unlearn (train)	Baseline (test)	After unlearn (test)
QA accuracy	0.84	0.32	0.88	0.32
QA fraction	1	0.119	1	0.111
SimDom accuracy	0.92	0.7	0.9	0.82
SimDom fraction	1	0.672	1	0.877
MMLU accuracy	0.62	0.6	0.65	0.632
MMLU fraction	1	0.946	1	0.955

---

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