diff --git a/logs/eval_IdkDPO_lr5e-05_beta0.5_alpha2_epoch5.log b/logs/eval_IdkDPO_lr5e-05_beta0.5_alpha2_epoch5.log
new file mode 100644
index 0000000000000000000000000000000000000000..310bcbc76fdfe91144dcfce5d1cdfb9e5d153dc5
--- /dev/null
+++ b/logs/eval_IdkDPO_lr5e-05_beta0.5_alpha2_epoch5.log
@@ -0,0 +1,78 @@
+/opt/conda/envs/copygen/lib/python3.12/site-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.6.3) or chardet (6.0.0.post1)/charset_normalizer (3.4.4) doesn't match a supported version!
+  warnings.warn(
+[2026-07-22 17:09:58,388][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/IdkDPO_lr5e-05_beta0.5_alpha2_epoch5
+[2026-07-22 17:09:58,389][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-22 17:09:58,389][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/IdkDPO_lr5e-05_beta0.5_alpha2_epoch5/TOFU_EVAL.json
+[2026-07-22 17:09:58,389][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/IdkDPO_lr5e-05_beta0.5_alpha2_epoch5/TOFU_SUMMARY.json
+[2026-07-22 17:09:59,059][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:04,  2.47it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:03,  3.34it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  3.68it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.91it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.09it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.18it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.27it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.38it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.41it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.41it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.39it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.37it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.56it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.19it/s]
+Evaluated 400 examples
+[2026-07-22 17:10:02,628][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:13,  1.12s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:12,  1.12s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:11,  1.16s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:10,  1.15s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:09,  1.14s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.13s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.13s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:09<00:05,  1.11s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:10<00:04,  1.11s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:11<00:03,  1.11s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:12<00:02,  1.12s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.12s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.06s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.11s/it]
+Evaluated 400 examples
+[2026-07-22 17:10:17,050][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-22 17:10:17,051][evaluator][INFO] - Result for metric forget_truth_ratio:	0.5424492803784302
+[2026-07-22 17:10:17,055][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-22 17:10:17,055][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:10:17,062][metrics][INFO] - Evaluating forget_quality
+[2026-07-22 17:10:17,062][evaluator][INFO] - Result for metric forget_quality:	1.0229984068189572e-13
+[2026-07-22 17:10:17,410][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.51it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.50it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.35it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:02,  4.33it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.37it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.43it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.42it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.52it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.55it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.54it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.51it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.49it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.72it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.52it/s]
+Evaluated 400 examples
+[2026-07-22 17:10:20,284][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.62306640625
+[2026-07-22 17:10:20,531][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:05,  2.28it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:04,  2.29it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:01<00:04,  2.41it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:03,  2.42it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:02<00:03,  2.39it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:02<00:02,  2.48it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:02<00:02,  2.71it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  3.11it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:03<00:01,  3.30it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:03<00:00,  3.12it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:03<00:00,  3.08it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:04<00:00,  2.93it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:04<00:00,  3.07it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:04<00:00,  2.82it/s]
+Evaluated 400 examples
+[2026-07-22 17:10:25,150][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.4722967646500736
+[2026-07-22 17:10:25,393][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.67it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.73it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.74it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.76it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.72it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.66it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.65it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.62it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.64it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.63it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.58it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.61it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.84it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.70it/s]
+Evaluated 400 examples
+[2026-07-22 17:10:28,408][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:04,  2.80it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:03,  2.82it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:03,  3.08it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:02,  3.03it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:02,  2.91it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:02<00:02,  2.38it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:02<00:02,  2.61it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  2.79it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:03<00:01,  2.99it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:03<00:01,  2.95it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:03<00:00,  2.79it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:04<00:00,  3.05it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:04<00:00,  3.15it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:04<00:00,  2.91it/s]
+Evaluated 400 examples
+[2026-07-22 17:10:33,302][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.63it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.64it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.61it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.66it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.69it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.58it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.54it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.49it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.48it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.50it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.47it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.50it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.74it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.60it/s]
+Evaluated 400 examples
+[2026-07-22 17:10:36,510][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:12,  1.07s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:11,  1.08s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:10,  1.08s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:09,  1.07s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:08,  1.06s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.07s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.09s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.09s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:09<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:10<00:03,  1.10s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:11<00:02,  1.11s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.10s/it]
Calculating loss: 100%|██████████| 13/13 [00:13<00:00,  1.05s/it]
Calculating loss: 100%|██████████| 13/13 [00:13<00:00,  1.08s/it]
+Evaluated 400 examples
+[2026-07-22 17:10:50,506][metrics][INFO] - Evaluating retain_Truth_Ratio
+[2026-07-22 17:10:50,839][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.07it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.53it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  7.46it/s]
+Evaluated 100 examples
+[2026-07-22 17:10:51,698][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.38it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.49it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.08it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.42it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.48it/s]
+Evaluated 100 examples
+[2026-07-22 17:10:53,311][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+[2026-07-22 17:10:53,617][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  5.77it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  4.91it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  4.22it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  4.88it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  4.81it/s]
+Evaluated 100 examples
+[2026-07-22 17:10:54,449][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-22 17:10:54,449][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:10:54,449][metrics][INFO] - Evaluating ra_Truth_Ratio
+[2026-07-22 17:10:54,699][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  7.71it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.34it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  8.02it/s]
+Evaluated 117 examples
+[2026-07-22 17:10:55,455][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.70it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.49it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.10it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.57it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.58it/s]
+Evaluated 117 examples
+[2026-07-22 17:10:57,007][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+[2026-07-22 17:10:57,307][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  4.82it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  4.46it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  4.05it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  4.59it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  4.49it/s]
+Evaluated 117 examples
+[2026-07-22 17:10:58,199][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-22 17:10:58,199][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:10:58,199][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-22 17:10:58,199][metrics][INFO] - Evaluating model_utility
+[2026-07-22 17:10:58,199][evaluator][INFO] - Result for metric model_utility:	0.5444555848679563
+[2026-07-22 17:10:58,733][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:10:58,740][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 14.21it/s]
 31%|███       | 4/13 [00:00<00:00, 12.90it/s]
 46%|████▌     | 6/13 [00:00<00:00, 13.33it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.77it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.98it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.70it/s]
100%|██████████| 13/13 [00:00<00:00, 14.24it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.49it/s]
 31%|███       | 4/13 [00:00<00:00, 13.32it/s]
 46%|████▌     | 6/13 [00:00<00:00, 13.70it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.73it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.82it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.92it/s]
100%|██████████| 13/13 [00:00<00:00, 14.24it/s]
+[2026-07-22 17:11:00,568][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:11:00,575][metrics][INFO] - Evaluating privleak
+[2026-07-22 17:11:00,575][evaluator][INFO] - Result for metric privleak:	-97.77995989588713
+[2026-07-22 17:11:00,839][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:00, 11.64it/s]
Calculating ES:  31%|███       | 4/13 [00:00<00:00, 10.44it/s]
Calculating ES:  46%|████▌     | 6/13 [00:00<00:00, 11.12it/s]
Calculating ES:  62%|██████▏   | 8/13 [00:00<00:00, 11.46it/s]
Calculating ES:  77%|███████▋  | 10/13 [00:00<00:00, 11.62it/s]
Calculating ES:  92%|█████████▏| 12/13 [00:01<00:00, 11.52it/s]
Calculating ES: 100%|██████████| 13/13 [00:01<00:00, 11.78it/s]
+Evaluated 400 examples
+[2026-07-22 17:11:01,943][evaluator][INFO] - Result for metric extraction_strength:	0.24642572663532428
diff --git a/logs/eval_IdkDPO_lr5e-05_beta0.5_alpha5_epoch10.log b/logs/eval_IdkDPO_lr5e-05_beta0.5_alpha5_epoch10.log
new file mode 100644
index 0000000000000000000000000000000000000000..377959babb4167d4ea5c8709bc50497ab7834cb9
--- /dev/null
+++ b/logs/eval_IdkDPO_lr5e-05_beta0.5_alpha5_epoch10.log
@@ -0,0 +1,76 @@
+[2026-07-21 07:49:18,817][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/IdkDPO_lr5e-05_beta0.5_alpha5_epoch10
+[2026-07-21 07:49:18,819][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-21 07:49:18,819][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/IdkDPO_lr5e-05_beta0.5_alpha5_epoch10/TOFU_EVAL.json
+[2026-07-21 07:49:18,819][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/IdkDPO_lr5e-05_beta0.5_alpha5_epoch10/TOFU_SUMMARY.json
+[2026-07-21 07:49:19,632][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:07,  1.51it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:04,  2.51it/s]
Calculating loss:  23%|██▎       | 3/13 [00:01<00:03,  3.16it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.62it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:02,  3.97it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.19it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.37it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:02<00:01,  4.54it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.62it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.68it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.69it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.69it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.85it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.10it/s]
+Evaluated 400 examples
+[2026-07-21 07:49:23,178][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:12,  1.02s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:11,  1.02s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:10,  1.04s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:09,  1.04s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:08,  1.03s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.03s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.03s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.01s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:09<00:04,  1.01s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:10<00:03,  1.01s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:11<00:02,  1.02s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:12<00:01,  1.02s/it]
Calculating loss: 100%|██████████| 13/13 [00:13<00:00,  1.03it/s]
Calculating loss: 100%|██████████| 13/13 [00:13<00:00,  1.01s/it]
+Evaluated 400 examples
+[2026-07-21 07:49:36,315][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-21 07:49:36,315][evaluator][INFO] - Result for metric forget_truth_ratio:	0.521592377972619
+[2026-07-21 07:49:36,323][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-21 07:49:36,325][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-21 07:49:36,340][metrics][INFO] - Evaluating forget_quality
+[2026-07-21 07:49:36,341][evaluator][INFO] - Result for metric forget_quality:	2.776495661774257e-16
+[2026-07-21 07:49:36,854][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.80it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.81it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.68it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.70it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.73it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.77it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.77it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.85it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.88it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.86it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.84it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.82it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  5.06it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.86it/s]
+Evaluated 400 examples
+[2026-07-21 07:49:39,529][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.6661669921875
+[2026-07-21 07:49:39,827][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:08,  1.43it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:01<00:08,  1.36it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:02<00:08,  1.24it/s]
Calculating text similarity:  31%|███       | 4/13 [00:02<00:06,  1.35it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:03<00:06,  1.31it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:04<00:05,  1.36it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:05<00:04,  1.46it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:05<00:03,  1.56it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:06<00:02,  1.50it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:06<00:02,  1.50it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:07<00:01,  1.56it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:08<00:00,  1.49it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:08<00:00,  1.61it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:08<00:00,  1.47it/s]
+Evaluated 400 examples
+[2026-07-21 07:49:48,649][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.5204367519573109
+[2026-07-21 07:49:49,021][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.94it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  5.02it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:01,  5.03it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  5.05it/s]
Calculating loss:  38%|███▊      | 5/13 [00:00<00:01,  5.05it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.97it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.96it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.92it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.96it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.95it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.90it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.94it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  5.18it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  5.02it/s]
+Evaluated 400 examples
+[2026-07-21 07:49:51,857][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:07,  1.56it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:01<00:07,  1.50it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:01<00:06,  1.64it/s]
Calculating text similarity:  31%|███       | 4/13 [00:02<00:05,  1.70it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:02<00:04,  1.75it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:03<00:04,  1.73it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:04<00:03,  1.65it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:04<00:03,  1.65it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:05<00:02,  1.64it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:06<00:01,  1.59it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:06<00:01,  1.53it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:07<00:00,  1.59it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:07<00:00,  1.63it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:07<00:00,  1.63it/s]
+Evaluated 400 examples
+[2026-07-21 07:50:00,188][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.92it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.93it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.91it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.96it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.97it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.89it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.87it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.82it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.81it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.83it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.80it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.82it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  5.08it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.92it/s]
+Evaluated 400 examples
+[2026-07-21 07:50:03,087][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:11,  1.02it/s]
Calculating loss:  15%|█▌        | 2/13 [00:01<00:10,  1.02it/s]
Calculating loss:  23%|██▎       | 3/13 [00:02<00:09,  1.01it/s]
Calculating loss:  31%|███       | 4/13 [00:03<00:08,  1.02it/s]
Calculating loss:  38%|███▊      | 5/13 [00:04<00:07,  1.02it/s]
Calculating loss:  46%|████▌     | 6/13 [00:05<00:06,  1.01it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:06<00:05,  1.00it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:07<00:05,  1.00s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:08<00:04,  1.01s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:09<00:03,  1.00s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:10<00:02,  1.01s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:11<00:01,  1.01s/it]
Calculating loss: 100%|██████████| 13/13 [00:12<00:00,  1.04it/s]
Calculating loss: 100%|██████████| 13/13 [00:12<00:00,  1.01it/s]
+Evaluated 400 examples
+[2026-07-21 07:50:15,929][metrics][INFO] - Evaluating retain_Truth_Ratio
+[2026-07-21 07:50:16,284][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.38it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.83it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  7.97it/s]
+Evaluated 100 examples
+[2026-07-21 07:50:17,038][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.31it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.63it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.22it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.64it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.67it/s]
+Evaluated 100 examples
+[2026-07-21 07:50:18,537][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+[2026-07-21 07:50:18,811][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:01,  2.63it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  2.71it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:01<00:00,  2.67it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  3.19it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  2.98it/s]
+Evaluated 100 examples
+[2026-07-21 07:50:20,154][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-21 07:50:20,154][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-21 07:50:20,154][metrics][INFO] - Evaluating ra_Truth_Ratio
+[2026-07-21 07:50:20,417][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.35it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.87it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  8.63it/s]
+Evaluated 117 examples
+[2026-07-21 07:50:21,130][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.49it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.61it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.21it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.70it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.71it/s]
+Evaluated 117 examples
+[2026-07-21 07:50:22,608][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+[2026-07-21 07:50:22,884][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:01,  2.45it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:01<00:01,  1.90it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:01<00:00,  1.95it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  2.05it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  2.04it/s]
+Evaluated 117 examples
+[2026-07-21 07:50:24,847][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-21 07:50:24,847][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-21 07:50:24,847][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-21 07:50:24,848][metrics][INFO] - Evaluating model_utility
+[2026-07-21 07:50:24,849][evaluator][INFO] - Result for metric model_utility:	0.5546318636120038
+[2026-07-21 07:50:25,425][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-21 07:50:25,441][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.05it/s]
 31%|███       | 4/13 [00:00<00:00, 11.27it/s]
 46%|████▌     | 6/13 [00:00<00:00, 11.55it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 11.70it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 11.95it/s]
 92%|█████████▏| 12/13 [00:01<00:00, 11.69it/s]
100%|██████████| 13/13 [00:01<00:00, 12.22it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 11.51it/s]
 31%|███       | 4/13 [00:00<00:00, 11.37it/s]
 46%|████▌     | 6/13 [00:00<00:00, 11.69it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 11.74it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 11.74it/s]
 92%|█████████▏| 12/13 [00:01<00:00, 11.85it/s]
100%|██████████| 13/13 [00:01<00:00, 12.14it/s]
+[2026-07-21 07:50:27,581][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-21 07:50:27,596][metrics][INFO] - Evaluating privleak
+[2026-07-21 07:50:27,596][evaluator][INFO] - Result for metric privleak:	-98.11504118594642
+[2026-07-21 07:50:27,882][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:01, 10.67it/s]
Calculating ES:  31%|███       | 4/13 [00:00<00:00,  9.49it/s]
Calculating ES:  46%|████▌     | 6/13 [00:00<00:00,  9.94it/s]
Calculating ES:  62%|██████▏   | 8/13 [00:00<00:00, 10.24it/s]
Calculating ES:  77%|███████▋  | 10/13 [00:00<00:00, 10.26it/s]
Calculating ES:  92%|█████████▏| 12/13 [00:01<00:00, 10.08it/s]
Calculating ES: 100%|██████████| 13/13 [00:01<00:00, 10.48it/s]
+Evaluated 400 examples
+[2026-07-21 07:50:29,123][evaluator][INFO] - Result for metric extraction_strength:	0.2952828489835585
diff --git a/logs/eval_IdkDPO_lr5e-05_beta0.5_alpha5_epoch5.log b/logs/eval_IdkDPO_lr5e-05_beta0.5_alpha5_epoch5.log
new file mode 100644
index 0000000000000000000000000000000000000000..f274b81024f3ab2c52cd8e22ff45afa0f920df61
--- /dev/null
+++ b/logs/eval_IdkDPO_lr5e-05_beta0.5_alpha5_epoch5.log
@@ -0,0 +1,78 @@
+/opt/conda/envs/copygen/lib/python3.12/site-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.6.3) or chardet (6.0.0.post1)/charset_normalizer (3.4.4) doesn't match a supported version!
+  warnings.warn(
+[2026-07-22 17:00:21,742][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/IdkDPO_lr5e-05_beta0.5_alpha5_epoch5
+[2026-07-22 17:00:21,743][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-22 17:00:21,743][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/IdkDPO_lr5e-05_beta0.5_alpha5_epoch5/TOFU_EVAL.json
+[2026-07-22 17:00:21,743][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/IdkDPO_lr5e-05_beta0.5_alpha5_epoch5/TOFU_SUMMARY.json
+[2026-07-22 17:00:22,484][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:05,  2.29it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:03,  3.19it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  3.58it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.84it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.05it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.16it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.26it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:02<00:01,  4.39it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.43it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.47it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.44it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.42it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.62it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.18it/s]
+Evaluated 400 examples
+[2026-07-22 17:00:25,906][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:13,  1.12s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:12,  1.12s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:11,  1.15s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:10,  1.15s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:09,  1.14s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.13s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.12s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.11s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:10<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:11<00:03,  1.10s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:12<00:02,  1.11s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.12s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.07s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.11s/it]
+Evaluated 400 examples
+[2026-07-22 17:00:40,312][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-22 17:00:40,313][evaluator][INFO] - Result for metric forget_truth_ratio:	0.5391518666563003
+[2026-07-22 17:00:40,317][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-22 17:00:40,317][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:00:40,324][metrics][INFO] - Evaluating forget_quality
+[2026-07-22 17:00:40,324][evaluator][INFO] - Result for metric forget_quality:	3.1508505165833036e-15
+[2026-07-22 17:00:40,815][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.50it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.50it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.34it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:02,  4.38it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.41it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.46it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.45it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.54it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.58it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.56it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.50it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.48it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.72it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.54it/s]
+Evaluated 400 examples
+[2026-07-22 17:00:43,681][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.6280224609375
+[2026-07-22 17:00:43,924][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:04,  2.60it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:04,  2.32it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:01<00:04,  2.24it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:03,  2.27it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:02<00:03,  2.18it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:02<00:02,  2.39it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:02<00:02,  2.60it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:03<00:01,  2.81it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:03<00:01,  2.88it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:03<00:01,  2.97it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:04<00:00,  3.04it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:04<00:00,  2.93it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:04<00:00,  2.91it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:04<00:00,  2.68it/s]
+Evaluated 400 examples
+[2026-07-22 17:00:48,781][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.4796897890575585
+[2026-07-22 17:00:49,027][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.66it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.72it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.73it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.75it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.75it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.68it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.67it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.63it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.67it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.65it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.59it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.59it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.81it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.70it/s]
+Evaluated 400 examples
+[2026-07-22 17:00:52,076][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:03,  3.03it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:03,  2.90it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:03,  3.21it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:02,  3.28it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:02,  3.10it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:01<00:02,  3.10it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:02<00:01,  3.09it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  3.08it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:02<00:01,  3.15it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:03<00:00,  3.05it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:03<00:00,  3.03it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:03<00:00,  2.84it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:04<00:00,  2.96it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:04<00:00,  3.04it/s]
+Evaluated 400 examples
+[2026-07-22 17:00:56,617][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.60it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.62it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.59it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.65it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.68it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.61it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.57it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.52it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.51it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.53it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.50it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.53it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.76it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.62it/s]
+Evaluated 400 examples
+[2026-07-22 17:00:59,670][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:12,  1.08s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:11,  1.07s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:10,  1.08s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:09,  1.08s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:08,  1.07s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.08s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.09s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.10s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:09<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:10<00:03,  1.10s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:12<00:02,  1.11s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.11s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.06s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.08s/it]
+Evaluated 400 examples
+[2026-07-22 17:01:13,744][metrics][INFO] - Evaluating retain_Truth_Ratio
+[2026-07-22 17:01:14,091][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.14it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.56it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  7.60it/s]
+Evaluated 100 examples
+[2026-07-22 17:01:14,870][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.45it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.51it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.09it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.47it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.51it/s]
+Evaluated 100 examples
+[2026-07-22 17:01:16,462][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+[2026-07-22 17:01:16,763][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  5.78it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  5.03it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  4.56it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  5.46it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  5.25it/s]
+Evaluated 100 examples
+[2026-07-22 17:01:17,525][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-22 17:01:17,525][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:01:17,525][metrics][INFO] - Evaluating ra_Truth_Ratio
+[2026-07-22 17:01:17,770][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.06it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.54it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  8.27it/s]
+Evaluated 117 examples
+[2026-07-22 17:01:18,498][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.57it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.47it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.06it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.52it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.54it/s]
+Evaluated 117 examples
+[2026-07-22 17:01:20,076][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+[2026-07-22 17:01:20,382][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  4.92it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  4.48it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  3.60it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  3.98it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  4.02it/s]
+Evaluated 117 examples
+[2026-07-22 17:01:21,376][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-22 17:01:21,377][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:01:21,377][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-22 17:01:21,377][metrics][INFO] - Evaluating model_utility
+[2026-07-22 17:01:21,377][evaluator][INFO] - Result for metric model_utility:	0.543886049725611
+[2026-07-22 17:01:21,890][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:01:21,899][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 14.04it/s]
 31%|███       | 4/13 [00:00<00:00, 12.63it/s]
 46%|████▌     | 6/13 [00:00<00:00, 13.19it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.50it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.83it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.54it/s]
100%|██████████| 13/13 [00:00<00:00, 14.05it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.20it/s]
 31%|███       | 4/13 [00:00<00:00, 13.02it/s]
 46%|████▌     | 6/13 [00:00<00:00, 13.45it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.44it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.52it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.64it/s]
100%|██████████| 13/13 [00:00<00:00, 13.95it/s]
+[2026-07-22 17:01:23,759][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:01:23,766][metrics][INFO] - Evaluating privleak
+[2026-07-22 17:01:23,766][evaluator][INFO] - Result for metric privleak:	-97.90852582893103
+[2026-07-22 17:01:24,030][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:00, 11.95it/s]
Calculating ES:  31%|███       | 4/13 [00:00<00:00, 10.56it/s]
Calculating ES:  46%|████▌     | 6/13 [00:00<00:00, 11.21it/s]
Calculating ES:  62%|██████▏   | 8/13 [00:00<00:00, 11.62it/s]
Calculating ES:  77%|███████▋  | 10/13 [00:00<00:00, 11.75it/s]
Calculating ES:  92%|█████████▏| 12/13 [00:01<00:00, 11.47it/s]
Calculating ES: 100%|██████████| 13/13 [00:01<00:00, 11.91it/s]
+Evaluated 400 examples
+[2026-07-22 17:01:25,122][evaluator][INFO] - Result for metric extraction_strength:	0.250188864524075
diff --git a/logs/eval_IdkNLL_lr1e-05_alpha10_epoch10.log b/logs/eval_IdkNLL_lr1e-05_alpha10_epoch10.log
new file mode 100644
index 0000000000000000000000000000000000000000..9882a1d60d0d51ba231c8b690fe63b8bde4a0e87
--- /dev/null
+++ b/logs/eval_IdkNLL_lr1e-05_alpha10_epoch10.log
@@ -0,0 +1,76 @@
+[2026-07-21 07:49:13,805][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/IdkNLL_lr1e-05_alpha10_epoch10
+[2026-07-21 07:49:13,806][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-21 07:49:13,806][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/IdkNLL_lr1e-05_alpha10_epoch10/TOFU_EVAL.json
+[2026-07-21 07:49:13,806][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/IdkNLL_lr1e-05_alpha10_epoch10/TOFU_SUMMARY.json
+[2026-07-21 07:49:14,461][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:07,  1.50it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:04,  2.51it/s]
Calculating loss:  23%|██▎       | 3/13 [00:01<00:03,  3.16it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.61it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:02,  3.94it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.17it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.36it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:02<00:01,  4.52it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.60it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.66it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.67it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.68it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.83it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.08it/s]
+Evaluated 400 examples
+[2026-07-21 07:49:17,978][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:12,  1.03s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:11,  1.02s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:10,  1.04s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:09,  1.04s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:08,  1.03s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.03s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.02s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.01s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:09<00:04,  1.01s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:10<00:03,  1.01s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:11<00:02,  1.02s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:12<00:01,  1.02s/it]
Calculating loss: 100%|██████████| 13/13 [00:13<00:00,  1.03it/s]
Calculating loss: 100%|██████████| 13/13 [00:13<00:00,  1.01s/it]
+Evaluated 400 examples
+[2026-07-21 07:49:31,092][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-21 07:49:31,093][evaluator][INFO] - Result for metric forget_truth_ratio:	0.47608075658606114
+[2026-07-21 07:49:31,101][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-21 07:49:31,102][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-21 07:49:31,118][metrics][INFO] - Evaluating forget_quality
+[2026-07-21 07:49:31,119][evaluator][INFO] - Result for metric forget_quality:	6.916893189690245e-21
+[2026-07-21 07:49:31,575][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.81it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.81it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.68it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.71it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.74it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.77it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.78it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.86it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.89it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.87it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.86it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.84it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  5.10it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.88it/s]
+Evaluated 400 examples
+[2026-07-21 07:49:34,241][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.842978515625
+[2026-07-21 07:49:34,509][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:07,  1.64it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:01<00:06,  1.60it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:02<00:07,  1.40it/s]
Calculating text similarity:  31%|███       | 4/13 [00:02<00:06,  1.38it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:03<00:05,  1.51it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:03<00:04,  1.69it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:04<00:04,  1.43it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:05<00:03,  1.57it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:05<00:02,  1.71it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:06<00:01,  1.74it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:06<00:01,  1.67it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:07<00:00,  1.68it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:07<00:00,  1.79it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:07<00:00,  1.63it/s]
+Evaluated 400 examples
+[2026-07-21 07:49:42,477][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.43768437843662367
+[2026-07-21 07:49:42,835][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.93it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  5.01it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:01,  5.02it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  5.04it/s]
Calculating loss:  38%|███▊      | 5/13 [00:00<00:01,  5.05it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.98it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.97it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.94it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.97it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.96it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.91it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.95it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  5.19it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  5.03it/s]
+Evaluated 400 examples
+[2026-07-21 07:49:45,737][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:06,  1.98it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:01<00:05,  1.91it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:01<00:05,  1.86it/s]
Calculating text similarity:  31%|███       | 4/13 [00:02<00:04,  1.96it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:02<00:04,  1.95it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:03<00:03,  1.88it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:03<00:03,  1.82it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:04<00:02,  1.78it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:04<00:02,  1.82it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:05<00:01,  1.85it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:06<00:01,  1.72it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:06<00:00,  1.63it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:07<00:00,  1.77it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:07<00:00,  1.81it/s]
+Evaluated 400 examples
+[2026-07-21 07:49:53,259][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.92it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.94it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.91it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.96it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.99it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.91it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.88it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.83it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.83it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.84it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.82it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.84it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  5.10it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.93it/s]
+Evaluated 400 examples
+[2026-07-21 07:49:56,166][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:11,  1.02it/s]
Calculating loss:  15%|█▌        | 2/13 [00:01<00:10,  1.02it/s]
Calculating loss:  23%|██▎       | 3/13 [00:02<00:09,  1.01it/s]
Calculating loss:  31%|███       | 4/13 [00:03<00:08,  1.02it/s]
Calculating loss:  38%|███▊      | 5/13 [00:04<00:07,  1.03it/s]
Calculating loss:  46%|████▌     | 6/13 [00:05<00:06,  1.01it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:06<00:05,  1.00it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:07<00:05,  1.00s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:08<00:04,  1.00s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:09<00:03,  1.00s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:10<00:02,  1.01s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:11<00:01,  1.01s/it]
Calculating loss: 100%|██████████| 13/13 [00:12<00:00,  1.04it/s]
Calculating loss: 100%|██████████| 13/13 [00:12<00:00,  1.01it/s]
+Evaluated 400 examples
+[2026-07-21 07:50:08,980][metrics][INFO] - Evaluating retain_Truth_Ratio
+[2026-07-21 07:50:09,342][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.43it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.85it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  8.00it/s]
+Evaluated 100 examples
+[2026-07-21 07:50:10,143][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.37it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.63it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.22it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.64it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.67it/s]
+Evaluated 100 examples
+[2026-07-21 07:50:11,640][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+[2026-07-21 07:50:11,903][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:01,  2.86it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  2.86it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:01<00:00,  2.97it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  2.96it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  2.94it/s]
+Evaluated 100 examples
+[2026-07-21 07:50:13,266][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-21 07:50:13,266][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-21 07:50:13,266][metrics][INFO] - Evaluating ra_Truth_Ratio
+[2026-07-21 07:50:13,517][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.37it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.88it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  8.65it/s]
+Evaluated 117 examples
+[2026-07-21 07:50:14,241][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.55it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.62it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.22it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.71it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.72it/s]
+Evaluated 117 examples
+[2026-07-21 07:50:15,711][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+[2026-07-21 07:50:15,977][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:01,  2.31it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  2.24it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:01<00:00,  2.49it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  2.41it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  2.39it/s]
+Evaluated 117 examples
+[2026-07-21 07:50:17,650][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-21 07:50:17,651][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-21 07:50:17,651][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-21 07:50:17,651][metrics][INFO] - Evaluating model_utility
+[2026-07-21 07:50:17,652][evaluator][INFO] - Result for metric model_utility:	0.5770912163051024
+[2026-07-21 07:50:18,238][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-21 07:50:18,254][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.54it/s]
 31%|███       | 4/13 [00:00<00:00, 11.58it/s]
 46%|████▌     | 6/13 [00:00<00:00, 11.85it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 12.01it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 12.24it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 12.00it/s]
100%|██████████| 13/13 [00:01<00:00, 12.55it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 11.81it/s]
 31%|███       | 4/13 [00:00<00:00, 11.66it/s]
 46%|████▌     | 6/13 [00:00<00:00, 11.95it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 11.98it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 11.95it/s]
 92%|█████████▏| 12/13 [00:01<00:00, 12.08it/s]
100%|██████████| 13/13 [00:01<00:00, 12.40it/s]
+[2026-07-21 07:50:20,342][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-21 07:50:20,357][metrics][INFO] - Evaluating privleak
+[2026-07-21 07:50:20,357][evaluator][INFO] - Result for metric privleak:	-99.44929236509499
+[2026-07-21 07:50:20,639][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:00, 11.93it/s]
Calculating ES:  31%|███       | 4/13 [00:00<00:00, 10.39it/s]
Calculating ES:  46%|████▌     | 6/13 [00:00<00:00, 10.87it/s]
Calculating ES:  62%|██████▏   | 8/13 [00:00<00:00, 11.18it/s]
Calculating ES:  77%|███████▋  | 10/13 [00:00<00:00, 11.17it/s]
Calculating ES:  92%|█████████▏| 12/13 [00:01<00:00, 11.03it/s]
Calculating ES: 100%|██████████| 13/13 [00:01<00:00, 11.47it/s]
+Evaluated 400 examples
+[2026-07-21 07:50:21,772][evaluator][INFO] - Result for metric extraction_strength:	0.6415381995913018
diff --git a/logs/eval_IdkNLL_lr1e-05_alpha10_epoch5.log b/logs/eval_IdkNLL_lr1e-05_alpha10_epoch5.log
new file mode 100644
index 0000000000000000000000000000000000000000..2d99c4f35b4a857c7f46862b3bc055aa343f9d8a
--- /dev/null
+++ b/logs/eval_IdkNLL_lr1e-05_alpha10_epoch5.log
@@ -0,0 +1,78 @@
+/opt/conda/envs/copygen/lib/python3.12/site-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.6.3) or chardet (6.0.0.post1)/charset_normalizer (3.4.4) doesn't match a supported version!
+  warnings.warn(
+[2026-07-22 17:11:33,008][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/IdkNLL_lr1e-05_alpha10_epoch5
+[2026-07-22 17:11:33,009][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-22 17:11:33,009][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/IdkNLL_lr1e-05_alpha10_epoch5/TOFU_EVAL.json
+[2026-07-22 17:11:33,009][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/IdkNLL_lr1e-05_alpha10_epoch5/TOFU_SUMMARY.json
+[2026-07-22 17:11:34,408][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:05,  2.28it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:03,  3.17it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  3.54it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.79it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:02,  3.99it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.10it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.20it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:02<00:01,  4.33it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.38it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.42it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.37it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.34it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.50it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.11it/s]
+Evaluated 400 examples
+[2026-07-22 17:11:37,919][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:13,  1.12s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:12,  1.11s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:11,  1.14s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:10,  1.14s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:09,  1.13s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.13s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.13s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.11s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:10<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:11<00:03,  1.10s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:12<00:02,  1.11s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.12s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.06s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.10s/it]
+Evaluated 400 examples
+[2026-07-22 17:11:52,284][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-22 17:11:52,284][evaluator][INFO] - Result for metric forget_truth_ratio:	0.4756530184094565
+[2026-07-22 17:11:52,289][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-22 17:11:52,289][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:11:52,296][metrics][INFO] - Evaluating forget_quality
+[2026-07-22 17:11:52,297][evaluator][INFO] - Result for metric forget_quality:	6.916893189690245e-21
+[2026-07-22 17:11:52,663][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.47it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.45it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.30it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:02,  4.34it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.35it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.36it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.34it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.43it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.46it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.48it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.48it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.48it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.73it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.49it/s]
+Evaluated 400 examples
+[2026-07-22 17:11:55,559][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.8422119140625
+[2026-07-22 17:11:56,043][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:03,  3.23it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:03,  3.12it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:01<00:03,  2.65it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:03,  2.63it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:02,  2.96it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:01<00:02,  3.31it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:02<00:01,  3.11it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  3.31it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:02<00:01,  3.34it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:03<00:00,  3.27it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:03<00:00,  3.19it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:03<00:00,  3.19it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:04<00:00,  3.40it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:04<00:00,  3.18it/s]
+Evaluated 400 examples
+[2026-07-22 17:12:00,136][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.44177991317558574
+[2026-07-22 17:12:00,384][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.63it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.68it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.69it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.72it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.72it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.65it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.63it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.54it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.55it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.53it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.49it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.55it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.79it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.65it/s]
+Evaluated 400 examples
+[2026-07-22 17:12:03,424][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:03,  3.28it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:03,  3.46it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:02,  3.59it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:02,  3.86it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:02,  3.84it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:01<00:01,  3.65it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:01<00:01,  3.55it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  3.50it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:02<00:01,  3.57it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:02<00:00,  3.62it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:03<00:00,  3.36it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:03<00:00,  3.18it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.45it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.51it/s]
+Evaluated 400 examples
+[2026-07-22 17:12:07,380][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.50it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.57it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.56it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.62it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.64it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.57it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.53it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.48it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.49it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.50it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.47it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.50it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.70it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.57it/s]
+Evaluated 400 examples
+[2026-07-22 17:12:10,466][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:12,  1.07s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:11,  1.06s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:10,  1.07s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:09,  1.06s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:08,  1.06s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.07s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.09s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.10s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:09<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:10<00:03,  1.10s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:11<00:02,  1.10s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.10s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.05s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.08s/it]
+Evaluated 400 examples
+[2026-07-22 17:12:24,469][metrics][INFO] - Evaluating retain_Truth_Ratio
+[2026-07-22 17:12:24,956][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.05it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.51it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  7.49it/s]
+Evaluated 100 examples
+[2026-07-22 17:12:25,769][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.35it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.48it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.07it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.06it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.22it/s]
+Evaluated 100 examples
+[2026-07-22 17:12:27,570][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+[2026-07-22 17:12:27,809][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  5.47it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  5.50it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  5.32it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  5.66it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  5.57it/s]
+Evaluated 100 examples
+[2026-07-22 17:12:28,528][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-22 17:12:28,528][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:12:28,528][metrics][INFO] - Evaluating ra_Truth_Ratio
+[2026-07-22 17:12:28,766][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.07it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.53it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  8.29it/s]
+Evaluated 117 examples
+[2026-07-22 17:12:29,520][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.76it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.49it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.09it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.56it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.58it/s]
+Evaluated 117 examples
+[2026-07-22 17:12:31,071][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+[2026-07-22 17:12:31,316][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  3.86it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  3.74it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  4.21it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  4.24it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  4.13it/s]
+Evaluated 117 examples
+[2026-07-22 17:12:32,285][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-22 17:12:32,285][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:12:32,285][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-22 17:12:32,285][metrics][INFO] - Evaluating model_utility
+[2026-07-22 17:12:32,286][evaluator][INFO] - Result for metric model_utility:	0.5760213222711852
+[2026-07-22 17:12:32,780][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:12:32,787][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.75it/s]
 31%|███       | 4/13 [00:00<00:00, 12.47it/s]
 46%|████▌     | 6/13 [00:00<00:00, 12.81it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.01it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.50it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.26it/s]
100%|██████████| 13/13 [00:00<00:00, 13.73it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.41it/s]
 31%|███       | 4/13 [00:00<00:00, 13.07it/s]
 46%|████▌     | 6/13 [00:00<00:00, 13.60it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.64it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.45it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.73it/s]
100%|██████████| 13/13 [00:00<00:00, 14.05it/s]
+[2026-07-22 17:12:34,661][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:12:34,669][metrics][INFO] - Evaluating privleak
+[2026-07-22 17:12:34,669][evaluator][INFO] - Result for metric privleak:	-99.47561294981261
+[2026-07-22 17:12:35,386][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:00, 12.15it/s]
Calculating ES:  31%|███       | 4/13 [00:00<00:00, 11.16it/s]
Calculating ES:  46%|████▌     | 6/13 [00:00<00:00, 11.99it/s]
Calculating ES:  62%|██████▏   | 8/13 [00:00<00:00, 12.57it/s]
Calculating ES:  77%|███████▋  | 10/13 [00:00<00:00, 12.88it/s]
Calculating ES:  92%|█████████▏| 12/13 [00:00<00:00, 12.76it/s]
Calculating ES: 100%|██████████| 13/13 [00:00<00:00, 13.01it/s]
+Evaluated 400 examples
+[2026-07-22 17:12:36,386][evaluator][INFO] - Result for metric extraction_strength:	0.634425163480305
diff --git a/logs/eval_IdkNLL_lr1e-05_alpha1_epoch10.log b/logs/eval_IdkNLL_lr1e-05_alpha1_epoch10.log
new file mode 100644
index 0000000000000000000000000000000000000000..a1d2b4cdb53da83717d56649ecea721d0f50f9e3
--- /dev/null
+++ b/logs/eval_IdkNLL_lr1e-05_alpha1_epoch10.log
@@ -0,0 +1,78 @@
+/opt/conda/envs/copygen/lib/python3.12/site-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.6.3) or chardet (6.0.0.post1)/charset_normalizer (3.4.4) doesn't match a supported version!
+  warnings.warn(
+[2026-07-20 14:43:52,451][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/IdkNLL_lr1e-05_alpha1_epoch10
+[2026-07-20 14:43:52,453][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-20 14:43:52,453][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/IdkNLL_lr1e-05_alpha1_epoch10/TOFU_EVAL.json
+[2026-07-20 14:43:52,453][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/IdkNLL_lr1e-05_alpha1_epoch10/TOFU_SUMMARY.json
+[2026-07-20 14:43:53,109][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:10,  1.12it/s]
Calculating loss:  15%|█▌        | 2/13 [00:01<00:06,  1.58it/s]
Calculating loss:  23%|██▎       | 3/13 [00:01<00:05,  1.99it/s]
Calculating loss:  31%|███       | 4/13 [00:02<00:04,  2.07it/s]
Calculating loss:  38%|███▊      | 5/13 [00:02<00:03,  2.17it/s]
Calculating loss:  46%|████▌     | 6/13 [00:02<00:03,  2.23it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:03<00:02,  2.29it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:03<00:01,  2.53it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:04<00:01,  2.50it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:04<00:01,  2.48it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:04<00:00,  2.44it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:05<00:00,  2.41it/s]
Calculating loss: 100%|██████████| 13/13 [00:05<00:00,  2.57it/s]
Calculating loss: 100%|██████████| 13/13 [00:05<00:00,  2.28it/s]
+Evaluated 400 examples
+[2026-07-20 14:43:59,268][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:02<00:24,  2.02s/it]
Calculating loss:  15%|█▌        | 2/13 [00:04<00:22,  2.01s/it]
Calculating loss:  23%|██▎       | 3/13 [00:06<00:21,  2.19s/it]
Calculating loss:  31%|███       | 4/13 [00:08<00:19,  2.12s/it]
Calculating loss:  38%|███▊      | 5/13 [00:10<00:16,  2.06s/it]
Calculating loss:  46%|████▌     | 6/13 [00:12<00:14,  2.03s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:14<00:11,  2.00s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:15<00:09,  1.84s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:17<00:06,  1.74s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:19<00:05,  1.80s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:20<00:03,  1.66s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:22<00:01,  1.60s/it]
Calculating loss: 100%|██████████| 13/13 [00:23<00:00,  1.50s/it]
Calculating loss: 100%|██████████| 13/13 [00:23<00:00,  1.80s/it]
+Evaluated 400 examples
+[2026-07-20 14:44:22,620][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-20 14:44:22,621][evaluator][INFO] - Result for metric forget_truth_ratio:	0.48020106904725524
+[2026-07-20 14:44:22,629][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-20 14:44:22,630][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-20 14:44:22,646][metrics][INFO] - Evaluating forget_quality
+[2026-07-20 14:44:22,647][evaluator][INFO] - Result for metric forget_quality:	4.353260441808186e-19
+[2026-07-20 14:44:22,997][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:04,  2.63it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:04,  2.50it/s]
Calculating loss:  23%|██▎       | 3/13 [00:01<00:04,  2.39it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:03,  2.43it/s]
Calculating loss:  38%|███▊      | 5/13 [00:02<00:03,  2.41it/s]
Calculating loss:  46%|████▌     | 6/13 [00:02<00:02,  2.49it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:02<00:02,  2.55it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:03<00:01,  3.02it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:03<00:01,  3.35it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:03<00:00,  3.33it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:03<00:00,  3.30it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:04<00:00,  3.43it/s]
Calculating loss: 100%|██████████| 13/13 [00:04<00:00,  3.56it/s]
Calculating loss: 100%|██████████| 13/13 [00:04<00:00,  2.98it/s]
+Evaluated 400 examples
+[2026-07-20 14:44:27,365][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.79087890625
+[2026-07-20 14:44:27,677][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:02,  4.24it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:04,  2.73it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:02,  3.46it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:02,  4.22it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:02,  3.03it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:01<00:02,  2.76it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:02<00:02,  2.71it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  3.01it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:03<00:01,  2.59it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:03<00:01,  2.46it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:03<00:00,  2.85it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:04<00:00,  2.53it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:04<00:00,  3.03it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:04<00:00,  2.91it/s]
+Evaluated 400 examples
+[2026-07-20 14:44:32,138][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.025900074852939204
+[2026-07-20 14:44:32,477][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:03,  3.83it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.47it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.09it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:02,  4.32it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.06it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.20it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  3.86it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.04it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:01,  3.88it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.07it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  3.87it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.06it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.21it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.09it/s]
+Evaluated 400 examples
+[2026-07-20 14:44:35,960][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:07,  1.51it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:01<00:06,  1.73it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:01<00:06,  1.55it/s]
Calculating text similarity:  31%|███       | 4/13 [00:02<00:06,  1.35it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:03<00:05,  1.39it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:04<00:05,  1.35it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:05<00:04,  1.20it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:06<00:04,  1.16it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:06<00:03,  1.23it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:07<00:02,  1.26it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:08<00:01,  1.13it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:09<00:00,  1.18it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:10<00:00,  1.25it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:10<00:00,  1.27it/s]
+Evaluated 400 examples
+[2026-07-20 14:44:46,560][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:04,  2.56it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:03,  3.57it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  3.77it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.14it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:02,  2.88it/s]
Calculating loss:  46%|████▌     | 6/13 [00:02<00:02,  2.70it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:02<00:02,  2.60it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:02<00:01,  2.75it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:03<00:01,  2.61it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:03<00:01,  2.56it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:04<00:00,  2.50it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:04<00:00,  2.49it/s]
Calculating loss: 100%|██████████| 13/13 [00:04<00:00,  2.83it/s]
Calculating loss: 100%|██████████| 13/13 [00:04<00:00,  2.77it/s]
+Evaluated 400 examples
+[2026-07-20 14:44:51,527][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:22,  1.88s/it]
Calculating loss:  15%|█▌        | 2/13 [00:03<00:21,  1.94s/it]
Calculating loss:  23%|██▎       | 3/13 [00:05<00:19,  1.97s/it]
Calculating loss:  31%|███       | 4/13 [00:07<00:17,  1.95s/it]
Calculating loss:  38%|███▊      | 5/13 [00:09<00:15,  1.92s/it]
Calculating loss:  46%|████▌     | 6/13 [00:11<00:13,  1.93s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:13<00:11,  1.94s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:15<00:09,  1.94s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:17<00:07,  1.95s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:19<00:05,  1.96s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:21<00:03,  1.92s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:23<00:01,  1.88s/it]
Calculating loss: 100%|██████████| 13/13 [00:24<00:00,  1.81s/it]
Calculating loss: 100%|██████████| 13/13 [00:24<00:00,  1.90s/it]
+Evaluated 400 examples
+[2026-07-20 14:45:16,246][metrics][INFO] - Evaluating retain_Truth_Ratio
+[2026-07-20 14:45:16,621][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  9.27it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  4.15it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  3.50it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  3.47it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  3.73it/s]
+Evaluated 100 examples
+[2026-07-20 14:45:17,988][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.19it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:01,  1.86it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  1.57it/s]
Calculating loss: 100%|██████████| 4/4 [00:02<00:00,  1.62it/s]
Calculating loss: 100%|██████████| 4/4 [00:02<00:00,  1.73it/s]
+Evaluated 100 examples
+[2026-07-20 14:45:20,296][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+[2026-07-20 14:45:20,567][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:01,  2.15it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  2.15it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:01<00:00,  2.17it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  2.76it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  2.50it/s]
+Evaluated 100 examples
+[2026-07-20 14:45:22,167][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-20 14:45:22,167][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-20 14:45:22,167][metrics][INFO] - Evaluating ra_Truth_Ratio
+[2026-07-20 14:45:22,568][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  5.67it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  5.03it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  6.15it/s]
+Evaluated 117 examples
+[2026-07-20 14:45:23,473][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  4.37it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:01,  1.85it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  1.54it/s]
Calculating loss: 100%|██████████| 4/4 [00:02<00:00,  1.92it/s]
Calculating loss: 100%|██████████| 4/4 [00:02<00:00,  1.91it/s]
+Evaluated 117 examples
+[2026-07-20 14:45:25,568][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+[2026-07-20 14:45:25,856][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:01,  1.93it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  2.15it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:01<00:00,  2.05it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  2.01it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  2.03it/s]
+Evaluated 117 examples
+[2026-07-20 14:45:27,828][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-20 14:45:27,828][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-20 14:45:27,828][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-20 14:45:27,829][metrics][INFO] - Evaluating model_utility
+[2026-07-20 14:45:27,830][evaluator][INFO] - Result for metric model_utility:	0.4584928946618976
+[2026-07-20 14:45:28,384][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-20 14:45:28,399][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
  8%|▊         | 1/13 [00:00<00:01,  9.72it/s]
 15%|█▌        | 2/13 [00:00<00:01,  8.29it/s]
 23%|██▎       | 3/13 [00:00<00:01,  7.12it/s]
 31%|███       | 4/13 [00:00<00:01,  6.95it/s]
 38%|███▊      | 5/13 [00:00<00:01,  7.52it/s]
 54%|█████▍    | 7/13 [00:00<00:00,  9.10it/s]
 69%|██████▉   | 9/13 [00:00<00:00, 10.40it/s]
 85%|████████▍ | 11/13 [00:01<00:00,  5.98it/s]
 92%|█████████▏| 12/13 [00:01<00:00,  5.27it/s]
100%|██████████| 13/13 [00:01<00:00,  5.55it/s]
100%|██████████| 13/13 [00:01<00:00,  6.55it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
  8%|▊         | 1/13 [00:00<00:01,  7.76it/s]
 15%|█▌        | 2/13 [00:00<00:02,  4.63it/s]
 23%|██▎       | 3/13 [00:00<00:02,  4.05it/s]
 31%|███       | 4/13 [00:00<00:02,  3.85it/s]
 38%|███▊      | 5/13 [00:01<00:02,  3.73it/s]
 46%|████▌     | 6/13 [00:01<00:01,  3.75it/s]
 54%|█████▍    | 7/13 [00:01<00:01,  3.73it/s]
 62%|██████▏   | 8/13 [00:02<00:01,  3.59it/s]
 69%|██████▉   | 9/13 [00:02<00:01,  3.61it/s]
 77%|███████▋  | 10/13 [00:02<00:00,  3.80it/s]
 92%|█████████▏| 12/13 [00:02<00:00,  5.76it/s]
100%|██████████| 13/13 [00:02<00:00,  4.67it/s]
+[2026-07-20 14:45:33,173][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-20 14:45:33,187][metrics][INFO] - Evaluating privleak
+[2026-07-20 14:45:33,188][evaluator][INFO] - Result for metric privleak:	-99.44929236509498
+[2026-07-20 14:45:33,474][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:   8%|▊         | 1/13 [00:00<00:02,  5.22it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:02,  5.25it/s]
Calculating ES:  23%|██▎       | 3/13 [00:00<00:02,  4.23it/s]
Calculating ES:  31%|███       | 4/13 [00:00<00:02,  3.94it/s]
Calculating ES:  38%|███▊      | 5/13 [00:01<00:01,  4.87it/s]
Calculating ES:  46%|████▌     | 6/13 [00:01<00:01,  4.98it/s]
Calculating ES:  54%|█████▍    | 7/13 [00:01<00:01,  4.85it/s]
Calculating ES:  62%|██████▏   | 8/13 [00:01<00:01,  4.84it/s]
Calculating ES:  69%|██████▉   | 9/13 [00:01<00:00,  5.23it/s]
Calculating ES:  77%|███████▋  | 10/13 [00:02<00:00,  4.47it/s]
Calculating ES:  85%|████████▍ | 11/13 [00:02<00:00,  4.42it/s]
Calculating ES:  92%|█████████▏| 12/13 [00:02<00:00,  4.46it/s]
Calculating ES: 100%|██████████| 13/13 [00:02<00:00,  5.28it/s]
Calculating ES: 100%|██████████| 13/13 [00:02<00:00,  4.81it/s]
+Evaluated 400 examples
+[2026-07-20 14:45:36,179][evaluator][INFO] - Result for metric extraction_strength:	0.6068313205387513
diff --git a/logs/eval_IdkNLL_lr1e-05_alpha1_epoch5.log b/logs/eval_IdkNLL_lr1e-05_alpha1_epoch5.log
new file mode 100644
index 0000000000000000000000000000000000000000..f7f951d859eb1b373bf578ff1f0cc959814df2db
--- /dev/null
+++ b/logs/eval_IdkNLL_lr1e-05_alpha1_epoch5.log
@@ -0,0 +1,78 @@
+/opt/conda/envs/copygen/lib/python3.12/site-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.6.3) or chardet (6.0.0.post1)/charset_normalizer (3.4.4) doesn't match a supported version!
+  warnings.warn(
+[2026-07-22 17:08:48,772][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/IdkNLL_lr1e-05_alpha1_epoch5
+[2026-07-22 17:08:48,773][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-22 17:08:48,773][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/IdkNLL_lr1e-05_alpha1_epoch5/TOFU_EVAL.json
+[2026-07-22 17:08:48,773][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/IdkNLL_lr1e-05_alpha1_epoch5/TOFU_SUMMARY.json
+[2026-07-22 17:08:49,420][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:05,  2.28it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:03,  3.16it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  3.56it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.83it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.03it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.15it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.25it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:02<00:01,  4.38it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.42it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.46it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.44it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.42it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.62it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.17it/s]
+Evaluated 400 examples
+[2026-07-22 17:08:52,865][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:13,  1.11s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:12,  1.11s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:11,  1.14s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:10,  1.14s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:09,  1.14s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.13s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.12s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.10s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:10<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:11<00:03,  1.10s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:12<00:02,  1.11s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.12s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.06s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.10s/it]
+Evaluated 400 examples
+[2026-07-22 17:09:07,215][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-22 17:09:07,215][evaluator][INFO] - Result for metric forget_truth_ratio:	0.4817956462414547
+[2026-07-22 17:09:07,219][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-22 17:09:07,219][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:09:07,226][metrics][INFO] - Evaluating forget_quality
+[2026-07-22 17:09:07,227][evaluator][INFO] - Result for metric forget_quality:	1.11924339764853e-19
+[2026-07-22 17:09:07,644][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.52it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.52it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.37it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:02,  4.40it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.44it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.48it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.48it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.57it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.61it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.60it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.56it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.54it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.74it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.57it/s]
+Evaluated 400 examples
+[2026-07-22 17:09:10,491][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.792119140625
+[2026-07-22 17:09:10,737][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:01,  8.76it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:01,  6.05it/s]
Calculating text similarity:  31%|███       | 4/13 [00:00<00:01,  8.23it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:00<00:01,  5.73it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:00<00:01,  5.83it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:01<00:00,  6.09it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:01<00:00,  6.72it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:01<00:00,  6.04it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:01<00:00,  6.33it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:01<00:00,  6.05it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:01<00:00,  6.55it/s]
+Evaluated 400 examples
+[2026-07-22 17:09:12,724][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.02485689709593503
+[2026-07-22 17:09:12,975][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.52it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.64it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.67it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.69it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.67it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.57it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.55it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.53it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.60it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.61it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.58it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.63it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.86it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.67it/s]
+Evaluated 400 examples
+[2026-07-22 17:09:16,017][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:02,  4.43it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:02,  4.43it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:02,  4.11it/s]
Calculating text similarity:  31%|███       | 4/13 [00:00<00:02,  4.03it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:01,  4.28it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:01<00:01,  4.15it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:01<00:01,  3.81it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  3.65it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:02<00:01,  3.71it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:02<00:00,  3.54it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:02<00:00,  3.18it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:03<00:00,  3.47it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.80it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.79it/s]
+Evaluated 400 examples
+[2026-07-22 17:09:19,932][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.61it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.61it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.57it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.63it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.66it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.58it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.54it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.49it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.49it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.51it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.46it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.45it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.68it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.57it/s]
+Evaluated 400 examples
+[2026-07-22 17:09:23,019][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:12,  1.06s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:11,  1.06s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:10,  1.07s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:09,  1.06s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:08,  1.06s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.07s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.09s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.10s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:09<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:10<00:03,  1.10s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:11<00:02,  1.10s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.10s/it]
Calculating loss: 100%|██████████| 13/13 [00:13<00:00,  1.05s/it]
Calculating loss: 100%|██████████| 13/13 [00:13<00:00,  1.08s/it]
+Evaluated 400 examples
+[2026-07-22 17:09:37,012][metrics][INFO] - Evaluating retain_Truth_Ratio
+[2026-07-22 17:09:37,443][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.06it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.51it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  7.47it/s]
+Evaluated 100 examples
+[2026-07-22 17:09:38,322][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.25it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.49it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.08it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.11it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.26it/s]
+Evaluated 100 examples
+[2026-07-22 17:09:40,095][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+[2026-07-22 17:09:40,428][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  5.35it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  5.11it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  5.14it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  5.99it/s]
+Evaluated 100 examples
+[2026-07-22 17:09:41,097][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-22 17:09:41,097][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:09:41,097][metrics][INFO] - Evaluating ra_Truth_Ratio
+[2026-07-22 17:09:41,431][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.16it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.64it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  8.41it/s]
+Evaluated 117 examples
+[2026-07-22 17:09:42,171][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.77it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.49it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.10it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.56it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.58it/s]
+Evaluated 117 examples
+[2026-07-22 17:09:43,721][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+[2026-07-22 17:09:44,071][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  4.20it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  4.56it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  4.20it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  4.19it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  4.24it/s]
+Evaluated 117 examples
+[2026-07-22 17:09:45,016][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-22 17:09:45,016][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:09:45,016][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-22 17:09:45,017][metrics][INFO] - Evaluating model_utility
+[2026-07-22 17:09:45,017][evaluator][INFO] - Result for metric model_utility:	0.44971118166631735
+[2026-07-22 17:09:45,560][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:09:45,568][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.47it/s]
 31%|███       | 4/13 [00:00<00:00, 12.14it/s]
 46%|████▌     | 6/13 [00:00<00:00, 12.56it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 12.82it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.20it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 12.82it/s]
100%|██████████| 13/13 [00:00<00:00, 13.36it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 12.65it/s]
 31%|███       | 4/13 [00:00<00:00, 12.45it/s]
 46%|████▌     | 6/13 [00:00<00:00, 12.78it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 12.71it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 12.86it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 12.87it/s]
100%|██████████| 13/13 [00:00<00:00, 13.13it/s]
+[2026-07-22 17:09:47,533][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:09:47,541][metrics][INFO] - Evaluating privleak
+[2026-07-22 17:09:47,541][evaluator][INFO] - Result for metric privleak:	-99.48877324217145
+[2026-07-22 17:09:48,020][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:00, 11.93it/s]
Calculating ES:  31%|███       | 4/13 [00:00<00:00, 10.65it/s]
Calculating ES:  46%|████▌     | 6/13 [00:00<00:00, 11.21it/s]
Calculating ES:  62%|██████▏   | 8/13 [00:00<00:00, 11.64it/s]
Calculating ES:  77%|███████▋  | 10/13 [00:00<00:00, 11.80it/s]
Calculating ES:  92%|█████████▏| 12/13 [00:01<00:00, 11.77it/s]
Calculating ES: 100%|██████████| 13/13 [00:01<00:00, 12.10it/s]
+Evaluated 400 examples
+[2026-07-22 17:09:49,095][evaluator][INFO] - Result for metric extraction_strength:	0.5879578746962085
diff --git a/logs/eval_IdkNLL_lr1e-05_alpha2_epoch10.log b/logs/eval_IdkNLL_lr1e-05_alpha2_epoch10.log
new file mode 100644
index 0000000000000000000000000000000000000000..63afe5fa40ee182e232b3bbfa374a15238b8bdc6
--- /dev/null
+++ b/logs/eval_IdkNLL_lr1e-05_alpha2_epoch10.log
@@ -0,0 +1,78 @@
+/opt/conda/envs/copygen/lib/python3.12/site-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.6.3) or chardet (6.0.0.post1)/charset_normalizer (3.4.4) doesn't match a supported version!
+  warnings.warn(
+[2026-07-22 17:02:01,512][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/IdkNLL_lr1e-05_alpha2_epoch10
+[2026-07-22 17:02:01,513][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-22 17:02:01,513][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/IdkNLL_lr1e-05_alpha2_epoch10/TOFU_EVAL.json
+[2026-07-22 17:02:01,513][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/IdkNLL_lr1e-05_alpha2_epoch10/TOFU_SUMMARY.json
+[2026-07-22 17:02:02,162][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:05,  2.25it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:03,  3.12it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  3.52it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.80it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.02it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.13it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.24it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:02<00:01,  4.36it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.41it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.45it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.43it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.41it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.62it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.15it/s]
+Evaluated 400 examples
+[2026-07-22 17:02:05,615][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:13,  1.11s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:12,  1.12s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:11,  1.15s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:10,  1.15s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:09,  1.14s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.13s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.12s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.10s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:10<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:11<00:03,  1.10s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:12<00:02,  1.11s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.12s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.07s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.11s/it]
+Evaluated 400 examples
+[2026-07-22 17:02:20,008][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-22 17:02:20,009][evaluator][INFO] - Result for metric forget_truth_ratio:	0.4754792238777897
+[2026-07-22 17:02:20,013][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-22 17:02:20,013][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:02:20,020][metrics][INFO] - Evaluating forget_quality
+[2026-07-22 17:02:20,020][evaluator][INFO] - Result for metric forget_quality:	2.8138576976529647e-20
+[2026-07-22 17:02:20,582][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.49it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.49it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.34it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:02,  4.38it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.41it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.46it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.46it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.55it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.58it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.57it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.54it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.52it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.72it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.55it/s]
+Evaluated 400 examples
+[2026-07-22 17:02:23,442][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.80720703125
+[2026-07-22 17:02:23,684][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:02,  5.64it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:02,  4.94it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:02,  4.38it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:02,  3.41it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:02,  3.54it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:01<00:01,  3.91it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:01<00:01,  4.27it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:01<00:01,  4.60it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:02<00:00,  4.82it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:02<00:00,  4.91it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:02<00:00,  3.92it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:02<00:00,  3.58it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  4.27it/s]
+Evaluated 400 examples
+[2026-07-22 17:02:26,731][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.07287559625204562
+[2026-07-22 17:02:26,997][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.40it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.52it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.61it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.67it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.70it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.65it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.66it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.63it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.67it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.65it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.59it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.63it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.85it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.68it/s]
+Evaluated 400 examples
+[2026-07-22 17:02:30,060][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:03,  3.68it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:02,  4.08it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:02,  3.77it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:02,  3.81it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:02,  3.79it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:01<00:01,  3.51it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:01<00:01,  3.37it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  3.33it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:02<00:01,  3.38it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:02<00:00,  3.35it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:03<00:00,  3.19it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:03<00:00,  3.04it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.28it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.40it/s]
+Evaluated 400 examples
+[2026-07-22 17:02:34,127][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.59it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.56it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.55it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.60it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.56it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.48it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.43it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.37it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.41it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.45it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.45it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.49it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.73it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.54it/s]
+Evaluated 400 examples
+[2026-07-22 17:02:37,275][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:12,  1.06s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:11,  1.06s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:10,  1.08s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:09,  1.07s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:08,  1.06s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.08s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.09s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.10s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:09<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:10<00:03,  1.10s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:11<00:02,  1.10s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.10s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.05s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.08s/it]
+Evaluated 400 examples
+[2026-07-22 17:02:51,287][metrics][INFO] - Evaluating retain_Truth_Ratio
+[2026-07-22 17:02:51,636][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.00it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.44it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  7.45it/s]
+Evaluated 100 examples
+[2026-07-22 17:02:52,412][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.53it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.52it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.10it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.50it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.54it/s]
+Evaluated 100 examples
+[2026-07-22 17:02:53,988][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+[2026-07-22 17:02:54,233][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  5.32it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  5.15it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  5.21it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  5.59it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  5.45it/s]
+Evaluated 100 examples
+[2026-07-22 17:02:54,968][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-22 17:02:54,968][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:02:54,968][metrics][INFO] - Evaluating ra_Truth_Ratio
+[2026-07-22 17:02:55,226][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.17it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.63it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  8.41it/s]
+Evaluated 117 examples
+[2026-07-22 17:02:55,945][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.67it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.48it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.09it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.56it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.57it/s]
+Evaluated 117 examples
+[2026-07-22 17:02:57,500][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+[2026-07-22 17:02:57,755][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  4.15it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  3.95it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  4.33it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  4.27it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  4.23it/s]
+Evaluated 117 examples
+[2026-07-22 17:02:58,712][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-22 17:02:58,712][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:02:58,712][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-22 17:02:58,712][metrics][INFO] - Evaluating model_utility
+[2026-07-22 17:02:58,713][evaluator][INFO] - Result for metric model_utility:	0.542546453505407
+[2026-07-22 17:02:59,304][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:02:59,316][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 12.73it/s]
 31%|███       | 4/13 [00:00<00:00, 11.46it/s]
 46%|████▌     | 6/13 [00:00<00:00, 12.17it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 12.62it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.08it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 12.57it/s]
100%|██████████| 13/13 [00:00<00:00, 13.06it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 12.83it/s]
 31%|███       | 4/13 [00:00<00:00, 11.95it/s]
 46%|████▌     | 6/13 [00:00<00:00, 12.05it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 11.25it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 11.35it/s]
 92%|█████████▏| 12/13 [00:01<00:00, 11.71it/s]
100%|██████████| 13/13 [00:01<00:00, 12.13it/s]
+[2026-07-22 17:03:01,385][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:03:01,392][metrics][INFO] - Evaluating privleak
+[2026-07-22 17:03:01,392][evaluator][INFO] - Result for metric privleak:	-99.36628129021624
+[2026-07-22 17:03:01,699][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:00, 12.94it/s]
Calculating ES:  31%|███       | 4/13 [00:00<00:00, 11.64it/s]
Calculating ES:  46%|████▌     | 6/13 [00:00<00:00, 11.68it/s]
Calculating ES:  62%|██████▏   | 8/13 [00:00<00:00, 12.02it/s]
Calculating ES:  77%|███████▋  | 10/13 [00:00<00:00, 12.30it/s]
Calculating ES:  92%|█████████▏| 12/13 [00:00<00:00, 12.28it/s]
Calculating ES: 100%|██████████| 13/13 [00:01<00:00, 12.68it/s]
+Evaluated 400 examples
+[2026-07-22 17:03:02,724][evaluator][INFO] - Result for metric extraction_strength:	0.6247114783743656
diff --git a/logs/eval_IdkNLL_lr1e-05_alpha2_epoch5.log b/logs/eval_IdkNLL_lr1e-05_alpha2_epoch5.log
new file mode 100644
index 0000000000000000000000000000000000000000..dc6f821bd15b03ba58c4866b5d198cb2aee4f7c5
--- /dev/null
+++ b/logs/eval_IdkNLL_lr1e-05_alpha2_epoch5.log
@@ -0,0 +1,78 @@
+/opt/conda/envs/copygen/lib/python3.12/site-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.6.3) or chardet (6.0.0.post1)/charset_normalizer (3.4.4) doesn't match a supported version!
+  warnings.warn(
+[2026-07-22 17:13:05,961][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/IdkNLL_lr1e-05_alpha2_epoch5
+[2026-07-22 17:13:05,962][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-22 17:13:05,962][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/IdkNLL_lr1e-05_alpha2_epoch5/TOFU_EVAL.json
+[2026-07-22 17:13:05,962][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/IdkNLL_lr1e-05_alpha2_epoch5/TOFU_SUMMARY.json
+[2026-07-22 17:13:11,704][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:04,  2.44it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:03,  3.32it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  3.67it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.91it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.09it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.19it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.26it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:02<00:01,  4.36it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.40it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.41it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.40it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.38it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.58it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.19it/s]
+Evaluated 400 examples
+[2026-07-22 17:13:15,136][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:13,  1.12s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:12,  1.12s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:11,  1.16s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:10,  1.15s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:09,  1.13s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.12s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.12s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.11s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:10<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:11<00:03,  1.11s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:12<00:02,  1.11s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.11s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.06s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.10s/it]
+Evaluated 400 examples
+[2026-07-22 17:13:29,499][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-22 17:13:29,499][evaluator][INFO] - Result for metric forget_truth_ratio:	0.4763998624487176
+[2026-07-22 17:13:29,503][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-22 17:13:29,504][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:13:29,511][metrics][INFO] - Evaluating forget_quality
+[2026-07-22 17:13:29,511][evaluator][INFO] - Result for metric forget_quality:	1.11924339764853e-19
+[2026-07-22 17:13:29,859][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.48it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.49it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.29it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:02,  4.33it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.37it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.42it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.42it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.51it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.55it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.53it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.51it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.49it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.73it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.52it/s]
+Evaluated 400 examples
+[2026-07-22 17:13:32,736][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.8068798828125
+[2026-07-22 17:13:32,981][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:02,  5.07it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:02,  4.83it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:02,  4.13it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:02,  3.11it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:02,  3.24it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:01<00:01,  3.70it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:01<00:01,  4.14it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:01<00:01,  4.60it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:02<00:00,  4.94it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:02<00:00,  5.17it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:02<00:00,  4.66it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:02<00:00,  4.77it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:02<00:00,  4.58it/s]
+Evaluated 400 examples
+[2026-07-22 17:13:35,821][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.06959552198172891
+[2026-07-22 17:13:36,101][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.67it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.74it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.75it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.77it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.77it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.70it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.69it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.65it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.69it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.67it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.61it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.61it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.83it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.72it/s]
+Evaluated 400 examples
+[2026-07-22 17:13:39,095][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:03,  3.39it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:02,  3.90it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:02,  3.72it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:02,  3.67it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:02,  3.61it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:01<00:02,  3.47it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:02<00:01,  3.29it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  3.20it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:02<00:01,  3.28it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:02<00:00,  3.21it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:03<00:00,  2.92it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:03<00:00,  2.97it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.22it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.29it/s]
+Evaluated 400 examples
+[2026-07-22 17:13:43,280][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.58it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.61it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.59it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.65it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.68it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.61it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.57it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.51it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.51it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.52it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.49it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.52it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.76it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.61it/s]
+Evaluated 400 examples
+[2026-07-22 17:13:46,374][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:12,  1.07s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:11,  1.07s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:10,  1.08s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:09,  1.07s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:08,  1.07s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.08s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.09s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.09s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:09<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:10<00:03,  1.09s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:11<00:02,  1.10s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.11s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.05s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.08s/it]
+Evaluated 400 examples
+[2026-07-22 17:14:00,402][metrics][INFO] - Evaluating retain_Truth_Ratio
+[2026-07-22 17:14:00,786][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.22it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.58it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  7.57it/s]
+Evaluated 100 examples
+[2026-07-22 17:14:01,639][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.46it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.50it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.10it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.50it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.54it/s]
+Evaluated 100 examples
+[2026-07-22 17:14:03,216][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+[2026-07-22 17:14:03,484][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  5.73it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  4.79it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  4.74it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  4.93it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  4.93it/s]
+Evaluated 100 examples
+[2026-07-22 17:14:04,296][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-22 17:14:04,296][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:14:04,296][metrics][INFO] - Evaluating ra_Truth_Ratio
+[2026-07-22 17:14:04,527][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  7.99it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.52it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  8.26it/s]
+Evaluated 117 examples
+[2026-07-22 17:14:05,251][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.62it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.49it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.08it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.55it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.56it/s]
+Evaluated 117 examples
+[2026-07-22 17:14:06,813][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+[2026-07-22 17:14:07,131][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  4.04it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  4.04it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  4.51it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  4.52it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  4.41it/s]
+Evaluated 117 examples
+[2026-07-22 17:14:08,039][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-22 17:14:08,039][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:14:08,039][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-22 17:14:08,039][metrics][INFO] - Evaluating model_utility
+[2026-07-22 17:14:08,039][evaluator][INFO] - Result for metric model_utility:	0.529183527132918
+[2026-07-22 17:14:08,559][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:14:08,566][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 14.21it/s]
 31%|███       | 4/13 [00:00<00:00, 12.95it/s]
 46%|████▌     | 6/13 [00:00<00:00, 13.58it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.95it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 14.33it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 14.04it/s]
100%|██████████| 13/13 [00:00<00:00, 14.52it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.73it/s]
 31%|███       | 4/13 [00:00<00:00, 13.47it/s]
 46%|████▌     | 6/13 [00:00<00:00, 13.92it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.94it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 14.00it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 14.08it/s]
100%|██████████| 13/13 [00:00<00:00, 14.43it/s]
+[2026-07-22 17:14:10,365][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:14:10,371][metrics][INFO] - Evaluating privleak
+[2026-07-22 17:14:10,372][evaluator][INFO] - Result for metric privleak:	-99.41588546910718
+[2026-07-22 17:14:10,623][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:00, 13.00it/s]
Calculating ES:  31%|███       | 4/13 [00:00<00:00, 11.74it/s]
Calculating ES:  46%|████▌     | 6/13 [00:00<00:00, 12.28it/s]
Calculating ES:  62%|██████▏   | 8/13 [00:00<00:00, 12.85it/s]
Calculating ES:  77%|███████▋  | 10/13 [00:00<00:00, 13.03it/s]
Calculating ES:  92%|█████████▏| 12/13 [00:00<00:00, 12.85it/s]
Calculating ES: 100%|██████████| 13/13 [00:00<00:00, 13.25it/s]
+Evaluated 400 examples
+[2026-07-22 17:14:11,605][evaluator][INFO] - Result for metric extraction_strength:	0.6153080531513968
diff --git a/logs/eval_IdkNLL_lr1e-05_alpha5_epoch10.log b/logs/eval_IdkNLL_lr1e-05_alpha5_epoch10.log
new file mode 100644
index 0000000000000000000000000000000000000000..d514a5d6b9bd98da0af9c77ad26552e40b3e6cb4
--- /dev/null
+++ b/logs/eval_IdkNLL_lr1e-05_alpha5_epoch10.log
@@ -0,0 +1,78 @@
+/opt/conda/envs/copygen/lib/python3.12/site-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.6.3) or chardet (6.0.0.post1)/charset_normalizer (3.4.4) doesn't match a supported version!
+  warnings.warn(
+[2026-07-22 17:10:14,173][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/IdkNLL_lr1e-05_alpha5_epoch10
+[2026-07-22 17:10:14,174][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-22 17:10:14,174][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/IdkNLL_lr1e-05_alpha5_epoch10/TOFU_EVAL.json
+[2026-07-22 17:10:14,174][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/IdkNLL_lr1e-05_alpha5_epoch10/TOFU_SUMMARY.json
+[2026-07-22 17:10:14,977][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:04,  2.46it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:03,  3.33it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  3.68it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.91it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.10it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.20it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.29it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.41it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.45it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.49it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.46it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.43it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.63it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.23it/s]
+Evaluated 400 examples
+[2026-07-22 17:10:18,373][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:13,  1.12s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:12,  1.12s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:11,  1.15s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:10,  1.15s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:09,  1.13s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.13s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.12s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.10s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:10<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:11<00:03,  1.10s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:12<00:02,  1.12s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.12s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.06s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.10s/it]
+Evaluated 400 examples
+[2026-07-22 17:10:32,731][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-22 17:10:32,732][evaluator][INFO] - Result for metric forget_truth_ratio:	0.47389271319887827
+[2026-07-22 17:10:32,736][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-22 17:10:32,736][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:10:32,743][metrics][INFO] - Evaluating forget_quality
+[2026-07-22 17:10:32,743][evaluator][INFO] - Result for metric forget_quality:	1.3990395281788573e-20
+[2026-07-22 17:10:33,189][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.52it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.51it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.36it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:02,  4.39it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.42it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.44it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.43it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.52it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.56it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.55it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.52it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.49it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.74it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.54it/s]
+Evaluated 400 examples
+[2026-07-22 17:10:36,053][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.82615234375
+[2026-07-22 17:10:36,306][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:03,  3.68it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:03,  3.13it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:01<00:03,  2.58it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:03,  2.48it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:02,  2.76it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:02<00:02,  3.05it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:02<00:01,  3.42it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  3.84it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:02<00:01,  3.95it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:03<00:00,  3.79it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:03<00:00,  3.43it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:03<00:00,  3.36it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.47it/s]
+Evaluated 400 examples
+[2026-07-22 17:10:40,048][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.2513266852163511
+[2026-07-22 17:10:40,360][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.68it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.74it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.76it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.78it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.78it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.71it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.70it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.66it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.69it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.68it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.58it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.62it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.84it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.73it/s]
+Evaluated 400 examples
+[2026-07-22 17:10:43,393][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:03,  3.16it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:03,  3.25it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:02,  3.36it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:02,  3.36it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:02,  3.38it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:01<00:02,  3.18it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:02<00:01,  3.14it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  3.10it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:02<00:01,  3.21it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:03<00:00,  3.16it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:03<00:00,  2.99it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:03<00:00,  2.90it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:04<00:00,  3.23it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:04<00:00,  3.18it/s]
+Evaluated 400 examples
+[2026-07-22 17:10:47,817][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.62it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.65it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.62it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.68it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.71it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.63it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.59it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.53it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.53it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.54it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.51it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.55it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.79it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.64it/s]
+Evaluated 400 examples
+[2026-07-22 17:10:50,875][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:12,  1.07s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:11,  1.06s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:10,  1.08s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:09,  1.08s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:08,  1.07s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.08s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.09s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.09s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:09<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:10<00:03,  1.09s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:11<00:02,  1.10s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.11s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.05s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.08s/it]
+Evaluated 400 examples
+[2026-07-22 17:11:04,894][metrics][INFO] - Evaluating retain_Truth_Ratio
+[2026-07-22 17:11:05,261][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.20it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.59it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  7.61it/s]
+Evaluated 100 examples
+[2026-07-22 17:11:06,027][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.44it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.50it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.10it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.51it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.54it/s]
+Evaluated 100 examples
+[2026-07-22 17:11:07,602][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+[2026-07-22 17:11:07,964][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  5.08it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  5.08it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  4.83it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  4.85it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  4.89it/s]
+Evaluated 100 examples
+[2026-07-22 17:11:08,782][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-22 17:11:08,782][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:11:08,782][metrics][INFO] - Evaluating ra_Truth_Ratio
+[2026-07-22 17:11:09,060][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.06it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.54it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  8.26it/s]
+Evaluated 117 examples
+[2026-07-22 17:11:10,126][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.47it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.47it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.08it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.53it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.55it/s]
+Evaluated 117 examples
+[2026-07-22 17:11:11,697][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+[2026-07-22 17:11:11,952][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  3.25it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  3.43it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  3.92it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  4.06it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  3.87it/s]
+Evaluated 117 examples
+[2026-07-22 17:11:12,985][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-22 17:11:12,985][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:11:12,986][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-22 17:11:12,986][metrics][INFO] - Evaluating model_utility
+[2026-07-22 17:11:12,986][evaluator][INFO] - Result for metric model_utility:	0.570631823217008
+[2026-07-22 17:11:13,513][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:11:13,520][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.90it/s]
 31%|███       | 4/13 [00:00<00:00, 12.49it/s]
 46%|████▌     | 6/13 [00:00<00:00, 13.06it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.35it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.62it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.28it/s]
100%|██████████| 13/13 [00:00<00:00, 13.82it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 12.89it/s]
 31%|███       | 4/13 [00:00<00:00, 12.76it/s]
 46%|████▌     | 6/13 [00:00<00:00, 13.06it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.14it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.13it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.28it/s]
100%|██████████| 13/13 [00:00<00:00, 13.60it/s]
+[2026-07-22 17:11:15,419][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:11:15,425][metrics][INFO] - Evaluating privleak
+[2026-07-22 17:11:15,425][evaluator][INFO] - Result for metric privleak:	-99.4158854691072
+[2026-07-22 17:11:15,682][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:00, 12.67it/s]
Calculating ES:  31%|███       | 4/13 [00:00<00:00, 11.39it/s]
Calculating ES:  46%|████▌     | 6/13 [00:00<00:00, 12.01it/s]
Calculating ES:  62%|██████▏   | 8/13 [00:00<00:00, 12.42it/s]
Calculating ES:  77%|███████▋  | 10/13 [00:00<00:00, 12.39it/s]
Calculating ES:  92%|█████████▏| 12/13 [00:00<00:00, 12.18it/s]
Calculating ES: 100%|██████████| 13/13 [00:01<00:00, 12.66it/s]
+Evaluated 400 examples
+[2026-07-22 17:11:16,709][evaluator][INFO] - Result for metric extraction_strength:	0.6355469275996612
diff --git a/logs/eval_IdkNLL_lr1e-05_alpha5_epoch5.log b/logs/eval_IdkNLL_lr1e-05_alpha5_epoch5.log
new file mode 100644
index 0000000000000000000000000000000000000000..1b85fcaddf833e1c13075d9dfa1780f3fd15dc2e
--- /dev/null
+++ b/logs/eval_IdkNLL_lr1e-05_alpha5_epoch5.log
@@ -0,0 +1,78 @@
+/opt/conda/envs/copygen/lib/python3.12/site-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.6.3) or chardet (6.0.0.post1)/charset_normalizer (3.4.4) doesn't match a supported version!
+  warnings.warn(
+[2026-07-22 17:03:40,894][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/IdkNLL_lr1e-05_alpha5_epoch5
+[2026-07-22 17:03:40,895][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-22 17:03:40,895][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/IdkNLL_lr1e-05_alpha5_epoch5/TOFU_EVAL.json
+[2026-07-22 17:03:40,895][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/IdkNLL_lr1e-05_alpha5_epoch5/TOFU_SUMMARY.json
+[2026-07-22 17:03:41,626][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:05,  2.11it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:03,  2.99it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  3.40it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.67it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:02,  3.90it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.05it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.18it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:02<00:01,  4.32it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.38it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.43it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.42it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.41it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.62it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.09it/s]
+Evaluated 400 examples
+[2026-07-22 17:03:45,117][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:13,  1.11s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:12,  1.12s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:11,  1.15s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:10,  1.15s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:09,  1.14s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.13s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.12s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.11s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:10<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:11<00:03,  1.10s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:12<00:02,  1.11s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.12s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.07s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.11s/it]
+Evaluated 400 examples
+[2026-07-22 17:03:59,518][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-22 17:03:59,518][evaluator][INFO] - Result for metric forget_truth_ratio:	0.47482963606711215
+[2026-07-22 17:03:59,523][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-22 17:03:59,523][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:03:59,530][metrics][INFO] - Evaluating forget_quality
+[2026-07-22 17:03:59,531][evaluator][INFO] - Result for metric forget_quality:	2.8138576976529647e-20
+[2026-07-22 17:03:59,935][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.48it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.48it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.34it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:02,  4.37it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.41it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.45it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.45it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.54it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.58it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.57it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.54it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.52it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.76it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.55it/s]
+Evaluated 400 examples
+[2026-07-22 17:04:02,791][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.8255029296875
+[2026-07-22 17:04:03,026][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:03,  3.11it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:03,  3.03it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:01<00:03,  2.62it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:03,  2.59it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:02,  2.90it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:02<00:02,  3.20it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:02<00:01,  3.55it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  3.95it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:02<00:01,  3.58it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:03<00:00,  3.44it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:03<00:00,  3.23it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:03<00:00,  3.16it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.38it/s]
+Evaluated 400 examples
+[2026-07-22 17:04:06,873][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.23682434548046824
+[2026-07-22 17:04:07,156][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.40it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.60it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.65it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.70it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.72it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.67it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.66it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.63it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.67it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.66it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.60it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.64it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.86it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.70it/s]
+Evaluated 400 examples
+[2026-07-22 17:04:10,220][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:03,  3.27it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:03,  3.66it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:02,  3.70it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:02,  3.69it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:02,  3.68it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:01<00:01,  3.58it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:01<00:01,  3.48it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  3.43it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:02<00:01,  3.51it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:02<00:00,  3.39it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:03<00:00,  3.02it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:03<00:00,  2.85it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.11it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.32it/s]
+Evaluated 400 examples
+[2026-07-22 17:04:14,499][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.59it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.60it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.50it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.54it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.56it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.50it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.49it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.46it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.47it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.50it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.48it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.51it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.75it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.57it/s]
+Evaluated 400 examples
+[2026-07-22 17:04:17,653][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:12,  1.06s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:11,  1.06s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:10,  1.08s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:09,  1.07s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:08,  1.06s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.08s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.09s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.10s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:09<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:10<00:03,  1.10s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:11<00:02,  1.11s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.10s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.06s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.08s/it]
+Evaluated 400 examples
+[2026-07-22 17:04:31,693][metrics][INFO] - Evaluating retain_Truth_Ratio
+[2026-07-22 17:04:32,039][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.06it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.52it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  7.52it/s]
+Evaluated 100 examples
+[2026-07-22 17:04:32,825][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.47it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.51it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.10it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.51it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.55it/s]
+Evaluated 100 examples
+[2026-07-22 17:04:34,395][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+[2026-07-22 17:04:34,636][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  6.28it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  5.49it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  4.93it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  5.42it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  5.39it/s]
+Evaluated 100 examples
+[2026-07-22 17:04:35,378][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-22 17:04:35,378][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:04:35,378][metrics][INFO] - Evaluating ra_Truth_Ratio
+[2026-07-22 17:04:35,630][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.08it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.50it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  8.26it/s]
+Evaluated 117 examples
+[2026-07-22 17:04:36,348][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.63it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.48it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.08it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.54it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.56it/s]
+Evaluated 117 examples
+[2026-07-22 17:04:37,911][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+[2026-07-22 17:04:38,159][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  3.85it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  3.78it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  4.02it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  4.04it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  3.99it/s]
+Evaluated 117 examples
+[2026-07-22 17:04:39,162][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-22 17:04:39,163][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:04:39,163][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-22 17:04:39,163][metrics][INFO] - Evaluating model_utility
+[2026-07-22 17:04:39,163][evaluator][INFO] - Result for metric model_utility:	0.567320633002962
+[2026-07-22 17:04:39,647][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:04:39,654][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.22it/s]
 31%|███       | 4/13 [00:00<00:00, 12.04it/s]
 46%|████▌     | 6/13 [00:00<00:00, 12.62it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.02it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.46it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.31it/s]
100%|██████████| 13/13 [00:00<00:00, 13.65it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 12.91it/s]
 31%|███       | 4/13 [00:00<00:00, 11.36it/s]
 46%|████▌     | 6/13 [00:00<00:00, 11.52it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 11.80it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 12.14it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 12.48it/s]
100%|██████████| 13/13 [00:01<00:00, 12.60it/s]
+[2026-07-22 17:04:41,641][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:04:41,648][metrics][INFO] - Evaluating privleak
+[2026-07-22 17:04:41,648][evaluator][INFO] - Result for metric privleak:	-99.40474983711124
+[2026-07-22 17:04:41,897][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:00, 11.92it/s]
Calculating ES:  31%|███       | 4/13 [00:00<00:01,  8.64it/s]
Calculating ES:  46%|████▌     | 6/13 [00:00<00:00,  9.55it/s]
Calculating ES:  62%|██████▏   | 8/13 [00:00<00:00, 10.44it/s]
Calculating ES:  77%|███████▋  | 10/13 [00:00<00:00, 11.12it/s]
Calculating ES:  92%|█████████▏| 12/13 [00:01<00:00, 11.44it/s]
Calculating ES: 100%|██████████| 13/13 [00:01<00:00, 11.27it/s]
+Evaluated 400 examples
+[2026-07-22 17:04:43,051][evaluator][INFO] - Result for metric extraction_strength:	0.6241283959217184
diff --git a/logs/eval_IdkNLL_lr2e-05_alpha10_epoch10.log b/logs/eval_IdkNLL_lr2e-05_alpha10_epoch10.log
new file mode 100644
index 0000000000000000000000000000000000000000..00cc3daec9b2669797b4b4469d0185a63e9ad3ee
--- /dev/null
+++ b/logs/eval_IdkNLL_lr2e-05_alpha10_epoch10.log
@@ -0,0 +1,76 @@
+[2026-07-21 07:49:18,023][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/IdkNLL_lr2e-05_alpha10_epoch10
+[2026-07-21 07:49:18,024][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-21 07:49:18,025][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/IdkNLL_lr2e-05_alpha10_epoch10/TOFU_EVAL.json
+[2026-07-21 07:49:18,025][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/IdkNLL_lr2e-05_alpha10_epoch10/TOFU_SUMMARY.json
+[2026-07-21 07:49:18,749][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:07,  1.50it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:04,  2.50it/s]
Calculating loss:  23%|██▎       | 3/13 [00:01<00:03,  3.14it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.59it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:02,  3.93it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.15it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.31it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:02<00:01,  4.49it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.58it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.64it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.66it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:03<00:00,  4.66it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.81it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.06it/s]
+Evaluated 400 examples
+[2026-07-21 07:49:22,356][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:12,  1.03s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:11,  1.03s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:10,  1.05s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:09,  1.05s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:08,  1.04s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.04s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.03s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.02s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:09<00:04,  1.02s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:10<00:03,  1.02s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:11<00:02,  1.02s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:12<00:01,  1.03s/it]
Calculating loss: 100%|██████████| 13/13 [00:13<00:00,  1.02it/s]
Calculating loss: 100%|██████████| 13/13 [00:13<00:00,  1.02s/it]
+Evaluated 400 examples
+[2026-07-21 07:49:35,564][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-21 07:49:35,565][evaluator][INFO] - Result for metric forget_truth_ratio:	0.4807089697996208
+[2026-07-21 07:49:35,573][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-21 07:49:35,574][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-21 07:49:35,589][metrics][INFO] - Evaluating forget_quality
+[2026-07-21 07:49:35,591][evaluator][INFO] - Result for metric forget_quality:	4.353260441808186e-19
+[2026-07-21 07:49:35,981][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.78it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.79it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.66it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.68it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.71it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.75it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.75it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.83it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.86it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.83it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.82it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.80it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  5.06it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.84it/s]
+Evaluated 400 examples
+[2026-07-21 07:49:38,665][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.797080078125
+[2026-07-21 07:49:38,948][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:06,  1.78it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:01<00:05,  2.03it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:01<00:05,  1.79it/s]
Calculating text similarity:  31%|███       | 4/13 [00:02<00:05,  1.78it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:02<00:04,  1.94it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:03<00:03,  2.02it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:03<00:02,  2.17it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:03<00:02,  2.15it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:04<00:01,  2.16it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:05<00:01,  1.99it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:05<00:01,  1.91it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:06<00:00,  1.83it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:06<00:00,  2.38it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:06<00:00,  2.06it/s]
+Evaluated 400 examples
+[2026-07-21 07:49:45,262][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.1882429679505736
+[2026-07-21 07:49:45,686][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.93it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  5.00it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.99it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  5.01it/s]
Calculating loss:  38%|███▊      | 5/13 [00:00<00:01,  5.01it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.94it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.93it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.90it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.93it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.92it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.88it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.91it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  5.15it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.99it/s]
+Evaluated 400 examples
+[2026-07-21 07:49:48,569][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:07,  1.59it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:01<00:06,  1.74it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:01<00:05,  1.77it/s]
Calculating text similarity:  31%|███       | 4/13 [00:02<00:04,  1.87it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:02<00:04,  1.96it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:03<00:03,  1.77it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:03<00:03,  1.68it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:04<00:02,  1.68it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:05<00:02,  1.68it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:05<00:01,  1.74it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:06<00:01,  1.54it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:07<00:00,  1.59it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:07<00:00,  1.73it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:07<00:00,  1.71it/s]
+Evaluated 400 examples
+[2026-07-21 07:49:56,565][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.89it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.90it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.87it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.93it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.96it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.88it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.85it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.80it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.80it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.81it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.79it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.81it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  5.07it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.90it/s]
+Evaluated 400 examples
+[2026-07-21 07:49:59,497][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:11,  1.01it/s]
Calculating loss:  15%|█▌        | 2/13 [00:01<00:10,  1.02it/s]
Calculating loss:  23%|██▎       | 3/13 [00:02<00:09,  1.01it/s]
Calculating loss:  31%|███       | 4/13 [00:03<00:08,  1.01it/s]
Calculating loss:  38%|███▊      | 5/13 [00:04<00:07,  1.02it/s]
Calculating loss:  46%|████▌     | 6/13 [00:05<00:06,  1.01it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:06<00:06,  1.00s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:07<00:05,  1.01s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:08<00:04,  1.01s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:09<00:03,  1.01s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:11<00:02,  1.01s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:12<00:01,  1.01s/it]
Calculating loss: 100%|██████████| 13/13 [00:12<00:00,  1.04it/s]
Calculating loss: 100%|██████████| 13/13 [00:12<00:00,  1.01it/s]
+Evaluated 400 examples
+[2026-07-21 07:50:12,391][metrics][INFO] - Evaluating retain_Truth_Ratio
+[2026-07-21 07:50:12,867][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.39it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.83it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  7.97it/s]
+Evaluated 100 examples
+[2026-07-21 07:50:13,686][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.27it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.61it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.21it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.63it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.66it/s]
+Evaluated 100 examples
+[2026-07-21 07:50:15,191][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+[2026-07-21 07:50:15,456][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  3.25it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  3.03it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:01<00:00,  2.76it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  2.64it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  2.74it/s]
+Evaluated 100 examples
+[2026-07-21 07:50:16,916][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-21 07:50:16,916][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-21 07:50:16,916][metrics][INFO] - Evaluating ra_Truth_Ratio
+[2026-07-21 07:50:17,219][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.31it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.85it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  8.60it/s]
+Evaluated 117 examples
+[2026-07-21 07:50:17,996][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.48it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.61it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.21it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.69it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.70it/s]
+Evaluated 117 examples
+[2026-07-21 07:50:19,476][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+[2026-07-21 07:50:19,733][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:01,  2.48it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  2.28it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:01<00:00,  2.34it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  2.33it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  2.33it/s]
+Evaluated 117 examples
+[2026-07-21 07:50:21,448][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-21 07:50:21,448][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-21 07:50:21,448][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-21 07:50:21,448][metrics][INFO] - Evaluating model_utility
+[2026-07-21 07:50:21,449][evaluator][INFO] - Result for metric model_utility:	0.57547938415897
+[2026-07-21 07:50:22,046][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-21 07:50:22,061][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 12.98it/s]
 31%|███       | 4/13 [00:00<00:00, 11.02it/s]
 46%|████▌     | 6/13 [00:00<00:00, 11.29it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 11.46it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 11.72it/s]
 92%|█████████▏| 12/13 [00:01<00:00, 11.47it/s]
100%|██████████| 13/13 [00:01<00:00, 11.99it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 11.25it/s]
 31%|███       | 4/13 [00:00<00:00, 11.11it/s]
 46%|████▌     | 6/13 [00:00<00:00, 11.41it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 11.44it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 11.45it/s]
 92%|█████████▏| 12/13 [00:01<00:00, 11.57it/s]
100%|██████████| 13/13 [00:01<00:00, 11.85it/s]
+[2026-07-21 07:50:24,247][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-21 07:50:24,261][metrics][INFO] - Evaluating privleak
+[2026-07-21 07:50:24,261][evaluator][INFO] - Result for metric privleak:	-99.17090156519676
+[2026-07-21 07:50:24,565][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:00, 11.41it/s]
Calculating ES:  31%|███       | 4/13 [00:00<00:00,  9.96it/s]
Calculating ES:  46%|████▌     | 6/13 [00:00<00:00, 10.39it/s]
Calculating ES:  62%|██████▏   | 8/13 [00:00<00:00, 10.63it/s]
Calculating ES:  77%|███████▋  | 10/13 [00:00<00:00, 10.66it/s]
Calculating ES:  92%|█████████▏| 12/13 [00:01<00:00, 10.50it/s]
Calculating ES: 100%|██████████| 13/13 [00:01<00:00, 10.94it/s]
+Evaluated 400 examples
+[2026-07-21 07:50:25,754][evaluator][INFO] - Result for metric extraction_strength:	0.5633441753353804
diff --git a/logs/eval_IdkNLL_lr2e-05_alpha10_epoch5.log b/logs/eval_IdkNLL_lr2e-05_alpha10_epoch5.log
new file mode 100644
index 0000000000000000000000000000000000000000..850ece3c3db9be69c8481d303c63e51f0be56875
--- /dev/null
+++ b/logs/eval_IdkNLL_lr2e-05_alpha10_epoch5.log
@@ -0,0 +1,78 @@
+/opt/conda/envs/copygen/lib/python3.12/site-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.6.3) or chardet (6.0.0.post1)/charset_normalizer (3.4.4) doesn't match a supported version!
+  warnings.warn(
+[2026-07-22 17:11:44,125][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/IdkNLL_lr2e-05_alpha10_epoch5
+[2026-07-22 17:11:44,126][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-22 17:11:44,126][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/IdkNLL_lr2e-05_alpha10_epoch5/TOFU_EVAL.json
+[2026-07-22 17:11:44,126][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/IdkNLL_lr2e-05_alpha10_epoch5/TOFU_SUMMARY.json
+[2026-07-22 17:11:44,903][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:05,  2.26it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:03,  3.16it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  3.51it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.75it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:02,  3.96it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.07it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.19it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:02<00:01,  4.33it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.40it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.45it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.44it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.42it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.63it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.14it/s]
+Evaluated 400 examples
+[2026-07-22 17:11:48,373][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:13,  1.11s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:12,  1.11s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:11,  1.14s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:10,  1.14s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:09,  1.13s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.13s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.12s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.10s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:10<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:11<00:03,  1.10s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:12<00:02,  1.11s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.11s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.06s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.10s/it]
+Evaluated 400 examples
+[2026-07-22 17:12:02,724][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-22 17:12:02,725][evaluator][INFO] - Result for metric forget_truth_ratio:	0.48126292587804714
+[2026-07-22 17:12:02,729][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-22 17:12:02,730][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:12:02,737][metrics][INFO] - Evaluating forget_quality
+[2026-07-22 17:12:02,737][evaluator][INFO] - Result for metric forget_quality:	4.353260441808186e-19
+[2026-07-22 17:12:03,083][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.41it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.47it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.34it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:02,  4.38it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.42it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.47it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.47it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.56it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.60it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.59it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.56it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.54it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.78it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.56it/s]
+Evaluated 400 examples
+[2026-07-22 17:12:05,932][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.7937744140625
+[2026-07-22 17:12:06,171][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:02,  4.26it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:03,  3.66it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:03,  2.80it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:03,  2.83it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:02,  3.21it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:01<00:02,  3.46it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:02<00:01,  3.64it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  3.79it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:02<00:01,  3.79it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:02<00:00,  3.68it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:03<00:00,  3.50it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:03<00:00,  3.33it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.62it/s]
+Evaluated 400 examples
+[2026-07-22 17:12:09,766][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.21226291796148353
+[2026-07-22 17:12:10,136][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.43it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.54it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.58it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.63it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.65it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.62it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.62it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.61it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.66it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.65it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.60it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.65it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.87it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.68it/s]
+Evaluated 400 examples
+[2026-07-22 17:12:13,138][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:03,  3.14it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:03,  3.50it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:02,  3.58it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:02,  3.80it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:02,  3.98it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:01<00:01,  3.81it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:01<00:01,  3.59it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  3.47it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:02<00:01,  3.52it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:02<00:00,  3.53it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:03<00:00,  3.31it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:03<00:00,  3.38it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.62it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.57it/s]
+Evaluated 400 examples
+[2026-07-22 17:12:17,049][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.60it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.58it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.52it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.60it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.63it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.56it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.49it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.42it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.41it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.43it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.41it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.46it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.72it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.55it/s]
+Evaluated 400 examples
+[2026-07-22 17:12:20,148][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:12,  1.06s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:11,  1.06s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:10,  1.07s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:09,  1.06s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:08,  1.06s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.08s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.10s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.10s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:09<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:10<00:03,  1.09s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:11<00:02,  1.10s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.10s/it]
Calculating loss: 100%|██████████| 13/13 [00:13<00:00,  1.05s/it]
Calculating loss: 100%|██████████| 13/13 [00:13<00:00,  1.08s/it]
+Evaluated 400 examples
+[2026-07-22 17:12:34,148][metrics][INFO] - Evaluating retain_Truth_Ratio
+[2026-07-22 17:12:35,333][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  7.89it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.37it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  6.55it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  6.68it/s]
+Evaluated 100 examples
+[2026-07-22 17:12:36,260][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.31it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.51it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.10it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.51it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.54it/s]
+Evaluated 100 examples
+[2026-07-22 17:12:37,835][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+[2026-07-22 17:12:38,127][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  5.69it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  4.43it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  4.56it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  4.94it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  4.85it/s]
+Evaluated 100 examples
+[2026-07-22 17:12:38,952][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-22 17:12:38,952][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:12:38,953][metrics][INFO] - Evaluating ra_Truth_Ratio
+[2026-07-22 17:12:39,186][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.08it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.59it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  8.35it/s]
+Evaluated 117 examples
+[2026-07-22 17:12:39,912][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.67it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.49it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.10it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.57it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.58it/s]
+Evaluated 117 examples
+[2026-07-22 17:12:41,462][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+[2026-07-22 17:12:41,731][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  4.43it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  4.07it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  3.97it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  4.08it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  4.08it/s]
+Evaluated 117 examples
+[2026-07-22 17:12:42,712][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-22 17:12:42,712][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:12:42,712][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-22 17:12:42,712][metrics][INFO] - Evaluating model_utility
+[2026-07-22 17:12:42,713][evaluator][INFO] - Result for metric model_utility:	0.5712395990961581
+[2026-07-22 17:12:43,297][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:12:43,305][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.74it/s]
 31%|███       | 4/13 [00:00<00:00, 12.30it/s]
 46%|████▌     | 6/13 [00:00<00:00, 12.58it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 12.79it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.16it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 12.88it/s]
100%|██████████| 13/13 [00:00<00:00, 13.40it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 12.58it/s]
 31%|███       | 4/13 [00:00<00:00, 12.49it/s]
 46%|████▌     | 6/13 [00:00<00:00, 12.96it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.02it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.08it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.28it/s]
100%|██████████| 13/13 [00:00<00:00, 13.54it/s]
+[2026-07-22 17:12:45,238][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:12:45,245][metrics][INFO] - Evaluating privleak
+[2026-07-22 17:12:45,245][evaluator][INFO] - Result for metric privleak:	-99.18709884809991
+[2026-07-22 17:12:45,523][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:00, 12.52it/s]
Calculating ES:  31%|███       | 4/13 [00:00<00:00, 11.31it/s]
Calculating ES:  46%|████▌     | 6/13 [00:00<00:00, 12.03it/s]
Calculating ES:  62%|██████▏   | 8/13 [00:00<00:00, 12.42it/s]
Calculating ES:  77%|███████▋  | 10/13 [00:00<00:00, 12.56it/s]
Calculating ES:  92%|█████████▏| 12/13 [00:00<00:00, 12.31it/s]
Calculating ES: 100%|██████████| 13/13 [00:01<00:00, 12.76it/s]
+Evaluated 400 examples
+[2026-07-22 17:12:46,542][evaluator][INFO] - Result for metric extraction_strength:	0.5530273167686603
diff --git a/logs/eval_IdkNLL_lr2e-05_alpha1_epoch10.log b/logs/eval_IdkNLL_lr2e-05_alpha1_epoch10.log
new file mode 100644
index 0000000000000000000000000000000000000000..be2ed46f47eb946f25076fc5d176f8dad9e105e8
--- /dev/null
+++ b/logs/eval_IdkNLL_lr2e-05_alpha1_epoch10.log
@@ -0,0 +1,78 @@
+/opt/conda/envs/copygen/lib/python3.12/site-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.6.3) or chardet (6.0.0.post1)/charset_normalizer (3.4.4) doesn't match a supported version!
+  warnings.warn(
+[2026-07-20 14:45:55,988][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/IdkNLL_lr2e-05_alpha1_epoch10
+[2026-07-20 14:45:55,989][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-20 14:45:55,989][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/IdkNLL_lr2e-05_alpha1_epoch10/TOFU_EVAL.json
+[2026-07-20 14:45:55,989][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/IdkNLL_lr2e-05_alpha1_epoch10/TOFU_SUMMARY.json
+[2026-07-20 14:45:56,677][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:10,  1.10it/s]
Calculating loss:  15%|█▌        | 2/13 [00:01<00:06,  1.61it/s]
Calculating loss:  23%|██▎       | 3/13 [00:01<00:05,  1.87it/s]
Calculating loss:  31%|███       | 4/13 [00:02<00:04,  2.13it/s]
Calculating loss:  38%|███▊      | 5/13 [00:02<00:03,  2.66it/s]
Calculating loss:  46%|████▌     | 6/13 [00:02<00:02,  2.90it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:03<00:02,  2.77it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:03<00:01,  2.69it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:03<00:01,  3.01it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:03<00:00,  3.39it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:04<00:00,  3.11it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:04<00:00,  2.84it/s]
Calculating loss: 100%|██████████| 13/13 [00:05<00:00,  2.80it/s]
Calculating loss: 100%|██████████| 13/13 [00:05<00:00,  2.58it/s]
+Evaluated 400 examples
+[2026-07-20 14:46:02,078][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:19,  1.66s/it]
Calculating loss:  15%|█▌        | 2/13 [00:03<00:16,  1.53s/it]
Calculating loss:  23%|██▎       | 3/13 [00:04<00:15,  1.50s/it]
Calculating loss:  31%|███       | 4/13 [00:06<00:13,  1.53s/it]
Calculating loss:  38%|███▊      | 5/13 [00:08<00:13,  1.69s/it]
Calculating loss:  46%|████▌     | 6/13 [00:09<00:10,  1.56s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:10<00:09,  1.53s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:12<00:07,  1.41s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:13<00:05,  1.40s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:14<00:04,  1.40s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:16<00:02,  1.43s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:17<00:01,  1.44s/it]
Calculating loss: 100%|██████████| 13/13 [00:18<00:00,  1.31s/it]
Calculating loss: 100%|██████████| 13/13 [00:18<00:00,  1.44s/it]
+Evaluated 400 examples
+[2026-07-20 14:46:20,861][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-20 14:46:20,862][evaluator][INFO] - Result for metric forget_truth_ratio:	0.4792364398260141
+[2026-07-20 14:46:20,870][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-20 14:46:20,871][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-20 14:46:20,886][metrics][INFO] - Evaluating forget_quality
+[2026-07-20 14:46:20,888][evaluator][INFO] - Result for metric forget_quality:	2.213522258546535e-19
+[2026-07-20 14:46:21,255][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:03,  3.53it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:03,  3.52it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:03,  3.31it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.41it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:02,  3.47it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  3.77it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  3.74it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:02<00:01,  4.00it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:01,  3.85it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  3.74it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  3.93it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:03<00:00,  3.99it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.39it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  3.86it/s]
+Evaluated 400 examples
+[2026-07-20 14:46:24,621][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.74496826171875
+[2026-07-20 14:46:24,981][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:02,  5.10it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:01,  5.99it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:01,  6.25it/s]
Calculating text similarity:  31%|███       | 4/13 [00:00<00:01,  6.45it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:02,  3.55it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:01<00:01,  4.26it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:01<00:01,  4.88it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:01<00:00,  5.31it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:01<00:00,  5.71it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:02<00:00,  4.89it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:02<00:00,  5.37it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:02<00:00,  5.75it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:02<00:00,  5.65it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:02<00:00,  5.24it/s]
+Evaluated 400 examples
+[2026-07-20 14:46:27,461][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.009941070158404315
+[2026-07-20 14:46:27,762][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:05,  2.33it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:04,  2.53it/s]
Calculating loss:  23%|██▎       | 3/13 [00:01<00:03,  2.62it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:03,  2.67it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:02,  2.68it/s]
Calculating loss:  46%|████▌     | 6/13 [00:02<00:02,  2.65it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:02<00:02,  2.66it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:03<00:01,  2.65it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:03<00:01,  2.68it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:03<00:01,  2.67it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:04<00:00,  2.59it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:04<00:00,  2.82it/s]
Calculating loss: 100%|██████████| 13/13 [00:04<00:00,  3.34it/s]
Calculating loss: 100%|██████████| 13/13 [00:04<00:00,  2.80it/s]
+Evaluated 400 examples
+[2026-07-20 14:46:32,692][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:11,  1.05it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:01<00:09,  1.14it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:02<00:09,  1.10it/s]
Calculating text similarity:  31%|███       | 4/13 [00:03<00:07,  1.19it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:04<00:07,  1.13it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:05<00:06,  1.16it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:06<00:05,  1.10it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:07<00:04,  1.03it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:08<00:03,  1.04it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:09<00:02,  1.02it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:10<00:02,  1.03s/it]
Calculating text similarity:  92%|█████████▏| 12/13 [00:11<00:01,  1.00s/it]
Calculating text similarity: 100%|██████████| 13/13 [00:12<00:00,  1.08it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:12<00:00,  1.07it/s]
+Evaluated 400 examples
+[2026-07-20 14:46:45,227][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:04,  2.46it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:04,  2.47it/s]
Calculating loss:  23%|██▎       | 3/13 [00:01<00:03,  2.78it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:03,  2.68it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:03,  2.60it/s]
Calculating loss:  46%|████▌     | 6/13 [00:02<00:02,  2.52it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:02<00:02,  2.49it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:03<00:01,  2.66it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:03<00:01,  2.58it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:03<00:01,  2.53it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:04<00:00,  2.48it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:04<00:00,  2.47it/s]
Calculating loss: 100%|██████████| 13/13 [00:04<00:00,  2.81it/s]
Calculating loss: 100%|██████████| 13/13 [00:04<00:00,  2.62it/s]
+Evaluated 400 examples
+[2026-07-20 14:46:50,494][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:23,  2.00s/it]
Calculating loss:  15%|█▌        | 2/13 [00:03<00:18,  1.69s/it]
Calculating loss:  23%|██▎       | 3/13 [00:05<00:17,  1.80s/it]
Calculating loss:  31%|███       | 4/13 [00:07<00:16,  1.83s/it]
Calculating loss:  38%|███▊      | 5/13 [00:08<00:13,  1.73s/it]
Calculating loss:  46%|████▌     | 6/13 [00:10<00:11,  1.64s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:11<00:09,  1.59s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:13<00:07,  1.54s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:14<00:06,  1.52s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:16<00:04,  1.50s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:17<00:03,  1.51s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:19<00:01,  1.49s/it]
Calculating loss: 100%|██████████| 13/13 [00:20<00:00,  1.42s/it]
Calculating loss: 100%|██████████| 13/13 [00:20<00:00,  1.57s/it]
+Evaluated 400 examples
+[2026-07-20 14:47:10,872][metrics][INFO] - Evaluating retain_Truth_Ratio
+[2026-07-20 14:47:11,281][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  5.65it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  4.94it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  4.78it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  4.92it/s]
+Evaluated 100 examples
+[2026-07-20 14:47:12,475][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.26it/s]
Calculating loss:  50%|█████     | 2/4 [00:01<00:01,  1.62it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:02<00:00,  1.24it/s]
Calculating loss: 100%|██████████| 4/4 [00:02<00:00,  1.21it/s]
Calculating loss: 100%|██████████| 4/4 [00:02<00:00,  1.34it/s]
+Evaluated 100 examples
+[2026-07-20 14:47:15,465][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+[2026-07-20 14:47:15,825][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:01,  2.22it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:01<00:01,  1.67it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:01<00:00,  1.93it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  2.49it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  2.22it/s]
+Evaluated 100 examples
+[2026-07-20 14:47:17,627][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-20 14:47:17,627][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-20 14:47:17,627][metrics][INFO] - Evaluating ra_Truth_Ratio
+[2026-07-20 14:47:17,973][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  5.65it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  4.74it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  6.08it/s]
+Evaluated 117 examples
+[2026-07-20 14:47:18,928][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  4.46it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:01,  1.86it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  1.58it/s]
Calculating loss: 100%|██████████| 4/4 [00:02<00:00,  1.90it/s]
Calculating loss: 100%|██████████| 4/4 [00:02<00:00,  1.91it/s]
+Evaluated 117 examples
+[2026-07-20 14:47:21,019][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+[2026-07-20 14:47:21,275][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:01,  2.14it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:01<00:01,  1.93it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:01<00:00,  2.09it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  2.02it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  2.03it/s]
+Evaluated 117 examples
+[2026-07-20 14:47:23,250][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-20 14:47:23,250][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-20 14:47:23,250][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-20 14:47:23,251][metrics][INFO] - Evaluating model_utility
+[2026-07-20 14:47:23,252][evaluator][INFO] - Result for metric model_utility:	0.4919533605980839
+[2026-07-20 14:47:23,921][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-20 14:47:23,936][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
  8%|▊         | 1/13 [00:00<00:01,  8.31it/s]
 15%|█▌        | 2/13 [00:00<00:01,  7.55it/s]
 23%|██▎       | 3/13 [00:00<00:01,  6.36it/s]
 31%|███       | 4/13 [00:00<00:01,  7.13it/s]
 38%|███▊      | 5/13 [00:00<00:01,  6.92it/s]
 46%|████▌     | 6/13 [00:00<00:00,  7.05it/s]
 54%|█████▍    | 7/13 [00:01<00:00,  6.80it/s]
 62%|██████▏   | 8/13 [00:01<00:00,  7.04it/s]
 69%|██████▉   | 9/13 [00:01<00:00,  7.12it/s]
 77%|███████▋  | 10/13 [00:01<00:00,  7.05it/s]
 85%|████████▍ | 11/13 [00:01<00:00,  7.62it/s]
 92%|█████████▏| 12/13 [00:01<00:00,  7.13it/s]
100%|██████████| 13/13 [00:01<00:00,  7.38it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
  8%|▊         | 1/13 [00:00<00:01,  6.75it/s]
 15%|█▌        | 2/13 [00:00<00:01,  6.72it/s]
 23%|██▎       | 3/13 [00:00<00:01,  6.64it/s]
 31%|███       | 4/13 [00:00<00:01,  7.07it/s]
 38%|███▊      | 5/13 [00:00<00:01,  7.01it/s]
 46%|████▌     | 6/13 [00:00<00:00,  7.22it/s]
 54%|█████▍    | 7/13 [00:00<00:00,  7.31it/s]
 62%|██████▏   | 8/13 [00:01<00:00,  7.38it/s]
 77%|███████▋  | 10/13 [00:01<00:00,  8.93it/s]
 92%|█████████▏| 12/13 [00:01<00:00,  9.98it/s]
100%|██████████| 13/13 [00:01<00:00,  8.64it/s]
+[2026-07-20 14:47:27,207][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-20 14:47:27,221][metrics][INFO] - Evaluating privleak
+[2026-07-20 14:47:27,221][evaluator][INFO] - Result for metric privleak:	-98.94110261400806
+[2026-07-20 14:47:27,588][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:   8%|▊         | 1/13 [00:00<00:03,  3.25it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:03,  3.43it/s]
Calculating ES:  23%|██▎       | 3/13 [00:00<00:02,  3.41it/s]
Calculating ES:  31%|███       | 4/13 [00:01<00:02,  3.42it/s]
Calculating ES:  38%|███▊      | 5/13 [00:01<00:02,  3.46it/s]
Calculating ES:  46%|████▌     | 6/13 [00:01<00:02,  3.49it/s]
Calculating ES:  54%|█████▍    | 7/13 [00:02<00:01,  3.47it/s]
Calculating ES:  62%|██████▏   | 8/13 [00:02<00:01,  3.54it/s]
Calculating ES:  69%|██████▉   | 9/13 [00:02<00:01,  3.53it/s]
Calculating ES:  77%|███████▋  | 10/13 [00:03<00:01,  2.98it/s]
Calculating ES:  92%|█████████▏| 12/13 [00:03<00:00,  4.52it/s]
Calculating ES: 100%|██████████| 13/13 [00:03<00:00,  3.98it/s]
+Evaluated 400 examples
+[2026-07-20 14:47:30,857][evaluator][INFO] - Result for metric extraction_strength:	0.5464293553274634
diff --git a/logs/eval_IdkNLL_lr2e-05_alpha1_epoch5.log b/logs/eval_IdkNLL_lr2e-05_alpha1_epoch5.log
new file mode 100644
index 0000000000000000000000000000000000000000..6f7c97565a5a4950fe5788cdde524a93aa243e70
--- /dev/null
+++ b/logs/eval_IdkNLL_lr2e-05_alpha1_epoch5.log
@@ -0,0 +1,78 @@
+/opt/conda/envs/copygen/lib/python3.12/site-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.6.3) or chardet (6.0.0.post1)/charset_normalizer (3.4.4) doesn't match a supported version!
+  warnings.warn(
+[2026-07-22 17:05:22,869][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/IdkNLL_lr2e-05_alpha1_epoch5
+[2026-07-22 17:05:22,870][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-22 17:05:22,870][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/IdkNLL_lr2e-05_alpha1_epoch5/TOFU_EVAL.json
+[2026-07-22 17:05:22,870][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/IdkNLL_lr2e-05_alpha1_epoch5/TOFU_SUMMARY.json
+[2026-07-22 17:05:23,525][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:04,  2.45it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:03,  3.32it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  3.67it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.90it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.08it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.18it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.27it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.39it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.43it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.46it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.44it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.42it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.58it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.20it/s]
+Evaluated 400 examples
+[2026-07-22 17:05:27,086][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:13,  1.12s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:12,  1.12s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:11,  1.15s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:10,  1.16s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:09,  1.14s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.13s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.12s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.11s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:10<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:11<00:03,  1.10s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:12<00:02,  1.12s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.12s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.06s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.11s/it]
+Evaluated 400 examples
+[2026-07-22 17:05:41,485][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-22 17:05:41,486][evaluator][INFO] - Result for metric forget_truth_ratio:	0.4781734859796574
+[2026-07-22 17:05:41,490][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-22 17:05:41,490][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:05:41,498][metrics][INFO] - Evaluating forget_quality
+[2026-07-22 17:05:41,498][evaluator][INFO] - Result for metric forget_quality:	1.11924339764853e-19
+[2026-07-22 17:05:41,839][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.50it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.50it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.35it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:02,  4.38it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.37it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.42it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.42it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.51it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.55it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.54it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.51it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.49it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.73it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.53it/s]
+Evaluated 400 examples
+[2026-07-22 17:05:44,712][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.74934814453125
+[2026-07-22 17:05:45,018][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:01,  8.61it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:01,  9.98it/s]
Calculating text similarity:  31%|███       | 4/13 [00:00<00:00,  9.55it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:00<00:00, 10.33it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:00<00:00, 10.64it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:01<00:00,  7.86it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:01<00:00,  8.65it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:01<00:00,  9.22it/s]
+Evaluated 400 examples
+[2026-07-22 17:05:46,429][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.006836801297299867
+[2026-07-22 17:05:46,678][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.37it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.52it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.60it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.67it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.70it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.65it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.66it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.63it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.66it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.65it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.60it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.64it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.87it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.69it/s]
+Evaluated 400 examples
+[2026-07-22 17:05:49,693][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:03,  3.24it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:02,  3.68it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:02,  3.70it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:02,  4.11it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:01,  4.09it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:01<00:01,  3.90it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:01<00:01,  3.68it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  3.26it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:02<00:01,  3.30it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:02<00:00,  3.33it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:03<00:00,  3.03it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:03<00:00,  3.02it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.27it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.42it/s]
+Evaluated 400 examples
+[2026-07-22 17:05:53,746][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.60it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.60it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.55it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.56it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.56it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.48it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.44it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.42it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.44it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.47it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.46it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.50it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.75it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.56it/s]
+Evaluated 400 examples
+[2026-07-22 17:05:56,839][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:12,  1.06s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:11,  1.06s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:10,  1.08s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:09,  1.07s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:08,  1.06s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.08s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.09s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.10s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:09<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:10<00:03,  1.10s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:11<00:02,  1.10s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.10s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.05s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.08s/it]
+Evaluated 400 examples
+[2026-07-22 17:06:10,849][metrics][INFO] - Evaluating retain_Truth_Ratio
+[2026-07-22 17:06:11,184][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  7.93it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.42it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  7.43it/s]
+Evaluated 100 examples
+[2026-07-22 17:06:11,988][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.56it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.52it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.11it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.52it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.56it/s]
+Evaluated 100 examples
+[2026-07-22 17:06:13,552][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+[2026-07-22 17:06:13,806][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  7.81it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  5.91it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  6.37it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  7.28it/s]
+Evaluated 100 examples
+[2026-07-22 17:06:14,357][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-22 17:06:14,357][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:06:14,357][metrics][INFO] - Evaluating ra_Truth_Ratio
+[2026-07-22 17:06:14,616][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.17it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.64it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  8.43it/s]
+Evaluated 117 examples
+[2026-07-22 17:06:15,327][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.76it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.48it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.09it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.56it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.57it/s]
+Evaluated 117 examples
+[2026-07-22 17:06:16,881][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+[2026-07-22 17:06:17,124][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  4.23it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  3.97it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  4.37it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  4.25it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  4.23it/s]
+Evaluated 117 examples
+[2026-07-22 17:06:18,070][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-22 17:06:18,070][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:06:18,070][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-22 17:06:18,071][metrics][INFO] - Evaluating model_utility
+[2026-07-22 17:06:18,071][evaluator][INFO] - Result for metric model_utility:	0.4494807077086565
+[2026-07-22 17:06:18,729][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:06:18,736][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.24it/s]
 31%|███       | 4/13 [00:00<00:00, 12.24it/s]
 46%|████▌     | 6/13 [00:00<00:00, 12.61it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 12.69it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.05it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 12.68it/s]
100%|██████████| 13/13 [00:00<00:00, 13.25it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 12.75it/s]
 31%|███       | 4/13 [00:00<00:00, 12.78it/s]
 46%|████▌     | 6/13 [00:00<00:00, 13.34it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.45it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.45it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.34it/s]
100%|██████████| 13/13 [00:00<00:00, 13.73it/s]
+[2026-07-22 17:06:20,666][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:06:20,672][metrics][INFO] - Evaluating privleak
+[2026-07-22 17:06:20,673][evaluator][INFO] - Result for metric privleak:	-99.03423699070129
+[2026-07-22 17:06:20,945][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:00, 12.56it/s]
Calculating ES:  31%|███       | 4/13 [00:00<00:00, 11.40it/s]
Calculating ES:  46%|████▌     | 6/13 [00:00<00:00, 12.13it/s]
Calculating ES:  62%|██████▏   | 8/13 [00:00<00:00, 12.55it/s]
Calculating ES:  77%|███████▋  | 10/13 [00:00<00:00, 12.72it/s]
Calculating ES:  92%|█████████▏| 12/13 [00:00<00:00, 12.50it/s]
Calculating ES: 100%|██████████| 13/13 [00:01<00:00, 12.93it/s]
+Evaluated 400 examples
+[2026-07-22 17:06:21,951][evaluator][INFO] - Result for metric extraction_strength:	0.536815840914721
diff --git a/logs/eval_IdkNLL_lr2e-05_alpha2_epoch10.log b/logs/eval_IdkNLL_lr2e-05_alpha2_epoch10.log
new file mode 100644
index 0000000000000000000000000000000000000000..46f4c8a239f568f1dc09563938d7f07f3410b1f4
--- /dev/null
+++ b/logs/eval_IdkNLL_lr2e-05_alpha2_epoch10.log
@@ -0,0 +1,78 @@
+/opt/conda/envs/copygen/lib/python3.12/site-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.6.3) or chardet (6.0.0.post1)/charset_normalizer (3.4.4) doesn't match a supported version!
+  warnings.warn(
+[2026-07-22 17:14:50,221][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/IdkNLL_lr2e-05_alpha2_epoch10
+[2026-07-22 17:14:50,222][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-22 17:14:50,222][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/IdkNLL_lr2e-05_alpha2_epoch10/TOFU_EVAL.json
+[2026-07-22 17:14:50,222][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/IdkNLL_lr2e-05_alpha2_epoch10/TOFU_SUMMARY.json
+[2026-07-22 17:14:50,922][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:04,  2.43it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:03,  3.30it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  3.65it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.89it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.08it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.18it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.27it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.39it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.43it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.47it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.45it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.43it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.59it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.20it/s]
+Evaluated 400 examples
+[2026-07-22 17:14:54,364][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:13,  1.12s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:12,  1.12s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:11,  1.15s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:10,  1.15s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:09,  1.13s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.13s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.12s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.11s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:10<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:11<00:03,  1.10s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:12<00:02,  1.12s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.12s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.06s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.11s/it]
+Evaluated 400 examples
+[2026-07-22 17:15:08,758][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-22 17:15:08,758][evaluator][INFO] - Result for metric forget_truth_ratio:	0.4787008684405464
+[2026-07-22 17:15:08,762][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-22 17:15:08,763][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:15:08,769][metrics][INFO] - Evaluating forget_quality
+[2026-07-22 17:15:08,770][evaluator][INFO] - Result for metric forget_quality:	1.11924339764853e-19
+[2026-07-22 17:15:09,179][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.51it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.50it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.35it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:02,  4.38it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.41it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.41it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.39it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.49it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.53it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.53it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.50it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.48it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.72it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.52it/s]
+Evaluated 400 examples
+[2026-07-22 17:15:12,055][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.75279541015625
+[2026-07-22 17:15:12,389][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:02,  5.38it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:01,  6.06it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:01,  7.35it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:00<00:01,  6.58it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:00<00:00,  8.05it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:01<00:00,  7.60it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:01<00:00,  7.18it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:01<00:00,  8.25it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:01<00:00,  7.84it/s]
+Evaluated 400 examples
+[2026-07-22 17:15:14,049][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.019885179327573877
+[2026-07-22 17:15:14,590][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.65it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.71it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.73it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.76it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.77it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.70it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.69it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.66it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.68it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.67it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.61it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.65it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.87it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.73it/s]
+Evaluated 400 examples
+[2026-07-22 17:15:17,619][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:03,  3.76it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:02,  3.93it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:02,  3.62it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:02,  3.68it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:02,  3.62it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:01<00:02,  3.50it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:01<00:01,  3.44it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  3.38it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:02<00:01,  3.46it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:02<00:00,  3.14it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:03<00:00,  2.94it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:03<00:00,  2.98it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.22it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.33it/s]
+Evaluated 400 examples
+[2026-07-22 17:15:21,801][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.57it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.50it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.46it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.51it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.55it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.51it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.50it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.47it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.49it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.52it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.50it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.53it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.77it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.57it/s]
+Evaluated 400 examples
+[2026-07-22 17:15:24,882][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:12,  1.06s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:11,  1.06s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:10,  1.08s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:09,  1.07s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:08,  1.06s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.08s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.09s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.09s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:09<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:10<00:03,  1.10s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:11<00:02,  1.10s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.10s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.06s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.08s/it]
+Evaluated 400 examples
+[2026-07-22 17:15:38,894][metrics][INFO] - Evaluating retain_Truth_Ratio
+[2026-07-22 17:15:39,283][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.08it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.53it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  7.51it/s]
+Evaluated 100 examples
+[2026-07-22 17:15:40,096][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.61it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.52it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.11it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.51it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.55it/s]
+Evaluated 100 examples
+[2026-07-22 17:15:41,667][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+[2026-07-22 17:15:41,933][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  6.04it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  5.41it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  5.02it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  6.13it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  5.81it/s]
+Evaluated 100 examples
+[2026-07-22 17:15:42,622][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-22 17:15:42,622][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:15:42,622][metrics][INFO] - Evaluating ra_Truth_Ratio
+[2026-07-22 17:15:42,856][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.14it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.54it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  8.31it/s]
+Evaluated 117 examples
+[2026-07-22 17:15:43,580][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.40it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.47it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.08it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.55it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.56it/s]
+Evaluated 117 examples
+[2026-07-22 17:15:45,142][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+[2026-07-22 17:15:45,388][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  4.06it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  3.87it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  4.08it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  4.07it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  4.04it/s]
+Evaluated 117 examples
+[2026-07-22 17:15:46,378][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-22 17:15:46,378][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:15:46,378][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-22 17:15:46,378][metrics][INFO] - Evaluating model_utility
+[2026-07-22 17:15:46,379][evaluator][INFO] - Result for metric model_utility:	0.5483017364641685
+[2026-07-22 17:15:46,897][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:15:46,905][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.26it/s]
 31%|███       | 4/13 [00:00<00:00, 12.24it/s]
 46%|████▌     | 6/13 [00:00<00:00, 12.95it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.33it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.74it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.44it/s]
100%|██████████| 13/13 [00:00<00:00, 13.89it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.78it/s]
 31%|███       | 4/13 [00:00<00:00, 13.62it/s]
 46%|████▌     | 6/13 [00:00<00:00, 14.05it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 14.07it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 14.13it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 14.24it/s]
100%|██████████| 13/13 [00:00<00:00, 14.58it/s]
+[2026-07-22 17:15:48,735][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:15:48,741][metrics][INFO] - Evaluating privleak
+[2026-07-22 17:15:48,741][evaluator][INFO] - Result for metric privleak:	-98.96033688745557
+[2026-07-22 17:15:49,108][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:00, 12.85it/s]
Calculating ES:  31%|███       | 4/13 [00:00<00:00, 11.58it/s]
Calculating ES:  46%|████▌     | 6/13 [00:00<00:00, 12.34it/s]
Calculating ES:  62%|██████▏   | 8/13 [00:00<00:00, 12.83it/s]
Calculating ES:  77%|███████▋  | 10/13 [00:00<00:00, 12.98it/s]
Calculating ES:  92%|█████████▏| 12/13 [00:00<00:00, 12.76it/s]
Calculating ES: 100%|██████████| 13/13 [00:00<00:00, 13.18it/s]
+Evaluated 400 examples
+[2026-07-22 17:15:50,094][evaluator][INFO] - Result for metric extraction_strength:	0.5469145524773721
diff --git a/logs/eval_IdkNLL_lr2e-05_alpha2_epoch5.log b/logs/eval_IdkNLL_lr2e-05_alpha2_epoch5.log
new file mode 100644
index 0000000000000000000000000000000000000000..9e2d2c55352123a5e3f3891bd59f96ad6c233f38
--- /dev/null
+++ b/logs/eval_IdkNLL_lr2e-05_alpha2_epoch5.log
@@ -0,0 +1,78 @@
+/opt/conda/envs/copygen/lib/python3.12/site-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.6.3) or chardet (6.0.0.post1)/charset_normalizer (3.4.4) doesn't match a supported version!
+  warnings.warn(
+[2026-07-22 17:13:16,964][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/IdkNLL_lr2e-05_alpha2_epoch5
+[2026-07-22 17:13:16,965][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-22 17:13:16,965][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/IdkNLL_lr2e-05_alpha2_epoch5/TOFU_EVAL.json
+[2026-07-22 17:13:16,965][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/IdkNLL_lr2e-05_alpha2_epoch5/TOFU_SUMMARY.json
+[2026-07-22 17:13:17,655][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:05,  2.28it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:03,  3.17it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  3.57it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.83it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.04it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.16it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.26it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:02<00:01,  4.39it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.44it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.48it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.46it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.44it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.65it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.18it/s]
+Evaluated 400 examples
+[2026-07-22 17:13:21,411][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:13,  1.11s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:12,  1.11s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:11,  1.14s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:10,  1.15s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:09,  1.14s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.13s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.12s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.10s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:10<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:11<00:03,  1.10s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:12<00:02,  1.11s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.12s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.06s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.10s/it]
+Evaluated 400 examples
+[2026-07-22 17:13:35,758][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-22 17:13:35,759][evaluator][INFO] - Result for metric forget_truth_ratio:	0.4780185789707035
+[2026-07-22 17:13:35,763][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-22 17:13:35,763][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:13:35,771][metrics][INFO] - Evaluating forget_quality
+[2026-07-22 17:13:35,771][evaluator][INFO] - Result for metric forget_quality:	1.3990395281788573e-20
+[2026-07-22 17:13:36,132][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.48it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.49it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.35it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:02,  4.39it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.42it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.47it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.47it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.56it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.60it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.59it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.52it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.50it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.74it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.55it/s]
+Evaluated 400 examples
+[2026-07-22 17:13:38,987][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.75410888671875
+[2026-07-22 17:13:39,231][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:02,  5.28it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:02,  4.91it/s]
Calculating text similarity:  31%|███       | 4/13 [00:00<00:01,  7.45it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:00<00:01,  5.89it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:01<00:00,  7.64it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:01<00:00,  7.43it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:01<00:00,  7.17it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:01<00:00,  7.53it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:01<00:00,  6.55it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:01<00:00,  7.05it/s]
+Evaluated 400 examples
+[2026-07-22 17:13:41,076][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.02179578252579108
+[2026-07-22 17:13:41,403][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.66it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.71it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.72it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.68it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.66it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.59it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.57it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.53it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.60it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.60it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.57it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.62it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.85it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.67it/s]
+Evaluated 400 examples
+[2026-07-22 17:13:44,433][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:03,  3.73it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:02,  3.85it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:02,  3.80it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:02,  3.63it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:02,  3.85it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:01<00:01,  3.54it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:01<00:01,  3.45it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  3.39it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:02<00:01,  3.44it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:02<00:00,  3.41it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:03<00:00,  3.25it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:03<00:00,  3.18it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.40it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.47it/s]
+Evaluated 400 examples
+[2026-07-22 17:13:48,501][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.61it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.62it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.59it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.65it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.66it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.59it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.54it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.48it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.45it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.44it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.40it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.43it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.70it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.57it/s]
+Evaluated 400 examples
+[2026-07-22 17:13:51,591][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:12,  1.06s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:11,  1.06s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:10,  1.07s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:09,  1.06s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:08,  1.06s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.07s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.09s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.10s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:09<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:10<00:03,  1.09s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:11<00:02,  1.10s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.10s/it]
Calculating loss: 100%|██████████| 13/13 [00:13<00:00,  1.05s/it]
Calculating loss: 100%|██████████| 13/13 [00:13<00:00,  1.08s/it]
+Evaluated 400 examples
+[2026-07-22 17:14:05,583][metrics][INFO] - Evaluating retain_Truth_Ratio
+[2026-07-22 17:14:05,948][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.08it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.52it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  7.47it/s]
+Evaluated 100 examples
+[2026-07-22 17:14:06,760][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.37it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.44it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.06it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.48it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.50it/s]
+Evaluated 100 examples
+[2026-07-22 17:14:08,358][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+[2026-07-22 17:14:08,590][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  6.32it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  4.61it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  4.70it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  5.76it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  5.42it/s]
+Evaluated 100 examples
+[2026-07-22 17:14:09,328][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-22 17:14:09,328][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:14:09,328][metrics][INFO] - Evaluating ra_Truth_Ratio
+[2026-07-22 17:14:09,619][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.19it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.64it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  8.42it/s]
+Evaluated 117 examples
+[2026-07-22 17:14:10,332][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.70it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.50it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.10it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.56it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.58it/s]
+Evaluated 117 examples
+[2026-07-22 17:14:11,885][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+[2026-07-22 17:14:12,133][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  4.60it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  4.25it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  4.25it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  4.36it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  4.35it/s]
+Evaluated 117 examples
+[2026-07-22 17:14:13,054][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-22 17:14:13,054][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:14:13,054][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-22 17:14:13,055][metrics][INFO] - Evaluating model_utility
+[2026-07-22 17:14:13,055][evaluator][INFO] - Result for metric model_utility:	0.5339999423611568
+[2026-07-22 17:14:13,634][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:14:13,641][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.55it/s]
 31%|███       | 4/13 [00:00<00:00, 12.20it/s]
 46%|████▌     | 6/13 [00:00<00:00, 12.66it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 12.75it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.20it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 12.80it/s]
100%|██████████| 13/13 [00:00<00:00, 13.37it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 12.85it/s]
 31%|███       | 4/13 [00:00<00:00, 12.39it/s]
 46%|████▌     | 6/13 [00:00<00:00, 12.93it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 12.99it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 12.78it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 12.94it/s]
100%|██████████| 13/13 [00:00<00:00, 13.28it/s]
+[2026-07-22 17:14:15,596][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:14:15,605][metrics][INFO] - Evaluating privleak
+[2026-07-22 17:14:15,605][evaluator][INFO] - Result for metric privleak:	-99.03929864160851
+[2026-07-22 17:14:15,873][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:00, 12.21it/s]
Calculating ES:  31%|███       | 4/13 [00:00<00:00, 10.90it/s]
Calculating ES:  46%|████▌     | 6/13 [00:00<00:00, 11.64it/s]
Calculating ES:  62%|██████▏   | 8/13 [00:00<00:00, 12.07it/s]
Calculating ES:  77%|███████▋  | 10/13 [00:00<00:00, 12.23it/s]
Calculating ES:  92%|█████████▏| 12/13 [00:01<00:00, 12.06it/s]
Calculating ES: 100%|██████████| 13/13 [00:01<00:00, 12.45it/s]
+Evaluated 400 examples
+[2026-07-22 17:14:16,918][evaluator][INFO] - Result for metric extraction_strength:	0.5295679579660575
diff --git a/logs/eval_IdkNLL_lr2e-05_alpha5_epoch10.log b/logs/eval_IdkNLL_lr2e-05_alpha5_epoch10.log
new file mode 100644
index 0000000000000000000000000000000000000000..9b82741e47ed60cda468cff3479c524bead9a9b3
--- /dev/null
+++ b/logs/eval_IdkNLL_lr2e-05_alpha5_epoch10.log
@@ -0,0 +1,78 @@
+/opt/conda/envs/copygen/lib/python3.12/site-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.6.3) or chardet (6.0.0.post1)/charset_normalizer (3.4.4) doesn't match a supported version!
+  warnings.warn(
+[2026-07-20 14:49:45,603][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/IdkNLL_lr2e-05_alpha5_epoch10
+[2026-07-20 14:49:45,604][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-20 14:49:45,604][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/IdkNLL_lr2e-05_alpha5_epoch10/TOFU_EVAL.json
+[2026-07-20 14:49:45,605][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/IdkNLL_lr2e-05_alpha5_epoch10/TOFU_SUMMARY.json
+[2026-07-20 14:49:46,497][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:08,  1.46it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:04,  2.46it/s]
Calculating loss:  23%|██▎       | 3/13 [00:01<00:03,  3.10it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.21it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:02,  3.28it/s]
Calculating loss:  46%|████▌     | 6/13 [00:02<00:02,  3.21it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:02<00:01,  3.25it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:02<00:01,  3.38it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:01,  3.45it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:03<00:00,  3.46it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:03<00:00,  3.50it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:03<00:00,  3.50it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  3.66it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  3.27it/s]
+Evaluated 400 examples
+[2026-07-20 14:49:50,979][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:15,  1.32s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:14,  1.31s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:12,  1.25s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:10,  1.21s/it]
Calculating loss:  38%|███▊      | 5/13 [00:06<00:09,  1.18s/it]
Calculating loss:  46%|████▌     | 6/13 [00:07<00:08,  1.17s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:08<00:06,  1.15s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:09<00:05,  1.12s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:10<00:04,  1.11s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:11<00:03,  1.12s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:12<00:02,  1.13s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.13s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.08s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.14s/it]
+Evaluated 400 examples
+[2026-07-20 14:50:05,866][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-20 14:50:05,867][evaluator][INFO] - Result for metric forget_truth_ratio:	0.48091575785024465
+[2026-07-20 14:50:05,875][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-20 14:50:05,877][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-20 14:50:05,892][metrics][INFO] - Evaluating forget_quality
+[2026-07-20 14:50:05,894][evaluator][INFO] - Result for metric forget_quality:	4.353260441808186e-19
+[2026-07-20 14:50:06,428][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.77it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.78it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.61it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.66it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.70it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.73it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.74it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.82it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.86it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.83it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.81it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.80it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  5.06it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.84it/s]
+Evaluated 400 examples
+[2026-07-20 14:50:09,117][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.775908203125
+[2026-07-20 14:50:09,601][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:05,  2.05it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:05,  2.17it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:01<00:04,  2.40it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:03,  2.62it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:02<00:03,  2.56it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:02<00:02,  2.69it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:02<00:02,  2.79it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  3.15it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:03<00:01,  3.06it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:03<00:00,  3.06it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:03<00:00,  2.99it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:04<00:00,  2.55it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:04<00:00,  3.19it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:04<00:00,  2.82it/s]
+Evaluated 400 examples
+[2026-07-20 14:50:14,209][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.06299447888449944
+[2026-07-20 14:50:14,612][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.86it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.94it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.93it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.97it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.99it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.93it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.89it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.86it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:01,  3.89it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  3.49it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  3.15it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:03<00:00,  3.02it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  3.05it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  3.80it/s]
+Evaluated 400 examples
+[2026-07-20 14:50:18,417][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:01<00:13,  1.11s/it]
Calculating text similarity:  15%|█▌        | 2/13 [00:01<00:10,  1.04it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:02<00:08,  1.25it/s]
Calculating text similarity:  31%|███       | 4/13 [00:03<00:06,  1.29it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:04<00:06,  1.25it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:05<00:06,  1.15it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:06<00:05,  1.04it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:07<00:04,  1.03it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:08<00:03,  1.01it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:09<00:03,  1.07s/it]
Calculating text similarity:  85%|████████▍ | 11/13 [00:10<00:02,  1.11s/it]
Calculating text similarity:  92%|█████████▏| 12/13 [00:11<00:01,  1.11s/it]
Calculating text similarity: 100%|██████████| 13/13 [00:12<00:00,  1.02s/it]
Calculating text similarity: 100%|██████████| 13/13 [00:12<00:00,  1.02it/s]
+Evaluated 400 examples
+[2026-07-20 14:50:36,612][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:04,  2.75it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:04,  2.58it/s]
Calculating loss:  23%|██▎       | 3/13 [00:01<00:03,  2.52it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:03,  2.52it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:02,  2.96it/s]
Calculating loss:  46%|████▌     | 6/13 [00:02<00:02,  3.40it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:02<00:01,  3.33it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:02<00:01,  3.01it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:03<00:01,  2.85it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:03<00:01,  2.77it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:03<00:00,  2.69it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:04<00:00,  2.59it/s]
Calculating loss: 100%|██████████| 13/13 [00:04<00:00,  2.72it/s]
Calculating loss: 100%|██████████| 13/13 [00:04<00:00,  2.80it/s]
+Evaluated 400 examples
+[2026-07-20 14:50:41,610][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:22,  1.88s/it]
Calculating loss:  15%|█▌        | 2/13 [00:03<00:16,  1.54s/it]
Calculating loss:  23%|██▎       | 3/13 [00:04<00:15,  1.54s/it]
Calculating loss:  31%|███       | 4/13 [00:06<00:13,  1.49s/it]
Calculating loss:  38%|███▊      | 5/13 [00:07<00:11,  1.47s/it]
Calculating loss:  46%|████▌     | 6/13 [00:09<00:10,  1.51s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:10<00:08,  1.49s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:12<00:07,  1.52s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:13<00:05,  1.49s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:15<00:04,  1.48s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:16<00:02,  1.48s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:17<00:01,  1.47s/it]
Calculating loss: 100%|██████████| 13/13 [00:19<00:00,  1.42s/it]
Calculating loss: 100%|██████████| 13/13 [00:19<00:00,  1.48s/it]
+Evaluated 400 examples
+[2026-07-20 14:51:00,901][metrics][INFO] - Evaluating retain_Truth_Ratio
+[2026-07-20 14:51:01,479][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  4.98it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  4.23it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  4.27it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  4.36it/s]
+Evaluated 100 examples
+[2026-07-20 14:51:02,780][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  4.42it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:01,  1.85it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  1.62it/s]
Calculating loss: 100%|██████████| 4/4 [00:02<00:00,  1.64it/s]
Calculating loss: 100%|██████████| 4/4 [00:02<00:00,  1.74it/s]
+Evaluated 100 examples
+[2026-07-20 14:51:05,078][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+[2026-07-20 14:51:05,392][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:01,  2.49it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  2.08it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:01<00:00,  2.16it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  2.60it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  2.43it/s]
+Evaluated 100 examples
+[2026-07-20 14:51:07,037][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-20 14:51:07,037][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-20 14:51:07,037][metrics][INFO] - Evaluating ra_Truth_Ratio
+[2026-07-20 14:51:07,322][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  5.61it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  4.59it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  5.81it/s]
+Evaluated 117 examples
+[2026-07-20 14:51:08,373][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:01,  2.55it/s]
Calculating loss:  50%|█████     | 2/4 [00:01<00:01,  1.72it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  1.57it/s]
Calculating loss: 100%|██████████| 4/4 [00:02<00:00,  1.85it/s]
Calculating loss: 100%|██████████| 4/4 [00:02<00:00,  1.82it/s]
+Evaluated 117 examples
+[2026-07-20 14:51:10,575][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+[2026-07-20 14:51:10,985][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:02,  1.42it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:01<00:01,  1.60it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:01<00:00,  1.53it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:02<00:00,  1.56it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:02<00:00,  1.55it/s]
+Evaluated 117 examples
+[2026-07-20 14:51:13,572][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-20 14:51:13,572][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-20 14:51:13,572][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-20 14:51:13,573][metrics][INFO] - Evaluating model_utility
+[2026-07-20 14:51:13,573][evaluator][INFO] - Result for metric model_utility:	0.5701453352863971
+[2026-07-20 14:51:14,178][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-20 14:51:14,195][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
  8%|▊         | 1/13 [00:00<00:01,  8.22it/s]
 15%|█▌        | 2/13 [00:00<00:01,  7.41it/s]
 23%|██▎       | 3/13 [00:00<00:01,  6.29it/s]
 31%|███       | 4/13 [00:00<00:01,  6.17it/s]
 38%|███▊      | 5/13 [00:00<00:01,  6.20it/s]
 46%|████▌     | 6/13 [00:00<00:01,  6.48it/s]
 54%|█████▍    | 7/13 [00:01<00:00,  6.27it/s]
 62%|██████▏   | 8/13 [00:01<00:00,  6.13it/s]
 69%|██████▉   | 9/13 [00:01<00:00,  6.47it/s]
 77%|███████▋  | 10/13 [00:01<00:00,  6.56it/s]
 85%|████████▍ | 11/13 [00:01<00:00,  6.48it/s]
 92%|█████████▏| 12/13 [00:01<00:00,  6.38it/s]
100%|██████████| 13/13 [00:01<00:00,  6.68it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
  8%|▊         | 1/13 [00:00<00:01,  6.01it/s]
 15%|█▌        | 2/13 [00:00<00:01,  6.09it/s]
 31%|███       | 4/13 [00:00<00:01,  7.95it/s]
 38%|███▊      | 5/13 [00:00<00:01,  7.47it/s]
 46%|████▌     | 6/13 [00:00<00:00,  7.49it/s]
 54%|█████▍    | 7/13 [00:00<00:00,  7.41it/s]
 62%|██████▏   | 8/13 [00:01<00:00,  7.09it/s]
 69%|██████▉   | 9/13 [00:01<00:00,  6.89it/s]
 77%|███████▋  | 10/13 [00:01<00:00,  6.99it/s]
 85%|████████▍ | 11/13 [00:01<00:00,  6.96it/s]
 92%|█████████▏| 12/13 [00:01<00:00,  6.87it/s]
100%|██████████| 13/13 [00:01<00:00,  7.28it/s]
+[2026-07-20 14:51:17,931][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-20 14:51:17,946][metrics][INFO] - Evaluating privleak
+[2026-07-20 14:51:17,946][evaluator][INFO] - Result for metric privleak:	-99.08687816013655
+[2026-07-20 14:51:18,239][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:   8%|▊         | 1/13 [00:00<00:03,  3.94it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:03,  3.29it/s]
Calculating ES:  23%|██▎       | 3/13 [00:00<00:03,  3.17it/s]
Calculating ES:  31%|███       | 4/13 [00:01<00:02,  4.33it/s]
Calculating ES:  38%|███▊      | 5/13 [00:01<00:01,  4.32it/s]
Calculating ES:  46%|████▌     | 6/13 [00:01<00:01,  4.51it/s]
Calculating ES:  54%|█████▍    | 7/13 [00:01<00:01,  4.19it/s]
Calculating ES:  62%|██████▏   | 8/13 [00:01<00:01,  4.24it/s]
Calculating ES:  69%|██████▉   | 9/13 [00:02<00:01,  3.91it/s]
Calculating ES:  77%|███████▋  | 10/13 [00:02<00:00,  3.58it/s]
Calculating ES:  85%|████████▍ | 11/13 [00:02<00:00,  3.24it/s]
Calculating ES:  92%|█████████▏| 12/13 [00:03<00:00,  3.24it/s]
Calculating ES: 100%|██████████| 13/13 [00:03<00:00,  3.77it/s]
Calculating ES: 100%|██████████| 13/13 [00:03<00:00,  3.78it/s]
+Evaluated 400 examples
+[2026-07-20 14:51:21,683][evaluator][INFO] - Result for metric extraction_strength:	0.5407521460937741
diff --git a/logs/eval_IdkNLL_lr2e-05_alpha5_epoch5.log b/logs/eval_IdkNLL_lr2e-05_alpha5_epoch5.log
new file mode 100644
index 0000000000000000000000000000000000000000..fb5c2df160ffb5312b4a343864c4c54370ef85e6
--- /dev/null
+++ b/logs/eval_IdkNLL_lr2e-05_alpha5_epoch5.log
@@ -0,0 +1,78 @@
+/opt/conda/envs/copygen/lib/python3.12/site-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.6.3) or chardet (6.0.0.post1)/charset_normalizer (3.4.4) doesn't match a supported version!
+  warnings.warn(
+[2026-07-22 17:06:51,713][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/IdkNLL_lr2e-05_alpha5_epoch5
+[2026-07-22 17:06:51,714][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-22 17:06:51,714][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/IdkNLL_lr2e-05_alpha5_epoch5/TOFU_EVAL.json
+[2026-07-22 17:06:51,714][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/IdkNLL_lr2e-05_alpha5_epoch5/TOFU_SUMMARY.json
+[2026-07-22 17:06:52,346][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:05,  2.23it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:03,  3.13it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  3.50it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.72it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:02,  3.92it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.04it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.16it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:02<00:01,  4.30it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.36it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.41it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.40it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.39it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.58it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.10it/s]
+Evaluated 400 examples
+[2026-07-22 17:06:55,832][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:13,  1.12s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:12,  1.12s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:11,  1.15s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:10,  1.15s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:09,  1.14s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.14s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.13s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:09<00:05,  1.11s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:10<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:11<00:03,  1.11s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:12<00:02,  1.12s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.12s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.07s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.11s/it]
+Evaluated 400 examples
+[2026-07-22 17:07:10,263][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-22 17:07:10,263][evaluator][INFO] - Result for metric forget_truth_ratio:	0.478923026212248
+[2026-07-22 17:07:10,268][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-22 17:07:10,269][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:07:10,277][metrics][INFO] - Evaluating forget_quality
+[2026-07-22 17:07:10,278][evaluator][INFO] - Result for metric forget_quality:	5.627715395305929e-20
+[2026-07-22 17:07:10,666][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.39it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.44it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.31it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:02,  4.36it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.40it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.45it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.45it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.55it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.59it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.58it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.55it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.53it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.77it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.55it/s]
+Evaluated 400 examples
+[2026-07-22 17:07:13,524][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.7750537109375
+[2026-07-22 17:07:13,767][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:02,  4.14it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:02,  4.14it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:02,  3.89it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:02,  3.83it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:01,  4.02it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:01<00:01,  4.09it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:01<00:01,  4.49it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:01<00:01,  4.68it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:02<00:00,  4.45it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:02<00:00,  5.00it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:02<00:00,  4.16it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:02<00:00,  3.88it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:02<00:00,  4.39it/s]
+Evaluated 400 examples
+[2026-07-22 17:07:16,727][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.09126959265861279
+[2026-07-22 17:07:17,012][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.59it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.68it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.69it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.71it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.72it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.59it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.56it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.53it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.58it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.59it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.56it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.61it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.84it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.67it/s]
+Evaluated 400 examples
+[2026-07-22 17:07:20,042][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:03,  3.62it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:02,  3.67it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:02,  3.72it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:02,  3.67it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:02,  3.72it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:01<00:01,  3.68it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:01<00:01,  3.57it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  3.47it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:02<00:01,  3.40it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:02<00:00,  3.43it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:03<00:00,  3.12it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:03<00:00,  3.06it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.36it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.44it/s]
+Evaluated 400 examples
+[2026-07-22 17:07:24,211][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.59it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.60it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.57it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.63it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.66it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.58it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.54it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.48it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.43it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.42it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.38it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.42it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.68it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.55it/s]
+Evaluated 400 examples
+[2026-07-22 17:07:27,304][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:12,  1.06s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:11,  1.06s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:10,  1.07s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:09,  1.07s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:08,  1.06s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.07s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.09s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.10s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:09<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:10<00:03,  1.10s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:11<00:02,  1.10s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.10s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.05s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.08s/it]
+Evaluated 400 examples
+[2026-07-22 17:07:41,313][metrics][INFO] - Evaluating retain_Truth_Ratio
+[2026-07-22 17:07:41,686][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.16it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.56it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  7.54it/s]
+Evaluated 100 examples
+[2026-07-22 17:07:42,457][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.31it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.46it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.08it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.50it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.52it/s]
+Evaluated 100 examples
+[2026-07-22 17:07:44,042][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+[2026-07-22 17:07:44,353][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  6.16it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  4.42it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  4.63it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  5.09it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  4.98it/s]
+Evaluated 100 examples
+[2026-07-22 17:07:45,157][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-22 17:07:45,157][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:07:45,157][metrics][INFO] - Evaluating ra_Truth_Ratio
+[2026-07-22 17:07:45,423][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.13it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.61it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  8.39it/s]
+Evaluated 117 examples
+[2026-07-22 17:07:46,193][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.75it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.50it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.09it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.55it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.57it/s]
+Evaluated 117 examples
+[2026-07-22 17:07:47,749][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+[2026-07-22 17:07:48,000][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  4.65it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  4.24it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  4.10it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  4.14it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  4.18it/s]
+Evaluated 117 examples
+[2026-07-22 17:07:48,959][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-22 17:07:48,959][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:07:48,959][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-22 17:07:48,959][metrics][INFO] - Evaluating model_utility
+[2026-07-22 17:07:48,959][evaluator][INFO] - Result for metric model_utility:	0.5644533474892907
+[2026-07-22 17:07:49,480][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:07:49,491][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.63it/s]
 31%|███       | 4/13 [00:00<00:00, 12.29it/s]
 46%|████▌     | 6/13 [00:00<00:00, 12.95it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 12.99it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.54it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.32it/s]
100%|██████████| 13/13 [00:00<00:00, 13.76it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.24it/s]
 31%|███       | 4/13 [00:00<00:00, 12.80it/s]
 46%|████▌     | 6/13 [00:00<00:00, 13.12it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.12it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.18it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.31it/s]
100%|██████████| 13/13 [00:00<00:00, 13.64it/s]
+[2026-07-22 17:07:51,391][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:07:51,398][metrics][INFO] - Evaluating privleak
+[2026-07-22 17:07:51,398][evaluator][INFO] - Result for metric privleak:	-99.10914942412843
+[2026-07-22 17:07:51,646][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:00, 12.68it/s]
Calculating ES:  31%|███       | 4/13 [00:00<00:00, 11.44it/s]
Calculating ES:  46%|████▌     | 6/13 [00:00<00:00, 12.20it/s]
Calculating ES:  62%|██████▏   | 8/13 [00:00<00:00, 12.64it/s]
Calculating ES:  77%|███████▋  | 10/13 [00:00<00:00, 12.17it/s]
Calculating ES:  92%|█████████▏| 12/13 [00:01<00:00, 11.69it/s]
Calculating ES: 100%|██████████| 13/13 [00:01<00:00, 12.44it/s]
+Evaluated 400 examples
+[2026-07-22 17:07:52,691][evaluator][INFO] - Result for metric extraction_strength:	0.545986699317964
diff --git a/logs/eval_IdkNLL_lr3e-05_alpha10_epoch10.log b/logs/eval_IdkNLL_lr3e-05_alpha10_epoch10.log
new file mode 100644
index 0000000000000000000000000000000000000000..9a97e0c8fc9e05ad3afe1a7c226aa59eb1e74e05
--- /dev/null
+++ b/logs/eval_IdkNLL_lr3e-05_alpha10_epoch10.log
@@ -0,0 +1,76 @@
+[2026-07-21 07:50:40,346][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/IdkNLL_lr3e-05_alpha10_epoch10
+[2026-07-21 07:50:40,348][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-21 07:50:40,348][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/IdkNLL_lr3e-05_alpha10_epoch10/TOFU_EVAL.json
+[2026-07-21 07:50:40,348][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/IdkNLL_lr3e-05_alpha10_epoch10/TOFU_SUMMARY.json
+[2026-07-21 07:50:41,087][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:07,  1.51it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:04,  2.51it/s]
Calculating loss:  23%|██▎       | 3/13 [00:01<00:03,  3.16it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.62it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:02,  3.97it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.20it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.38it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:02<00:01,  4.53it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.61it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.68it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.70it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.70it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.85it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.10it/s]
+Evaluated 400 examples
+[2026-07-21 07:50:44,609][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:12,  1.02s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:11,  1.02s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:10,  1.04s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:09,  1.04s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:08,  1.03s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.03s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.02s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.01s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:09<00:04,  1.01s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:10<00:03,  1.01s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:11<00:02,  1.02s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:12<00:01,  1.02s/it]
Calculating loss: 100%|██████████| 13/13 [00:13<00:00,  1.03it/s]
Calculating loss: 100%|██████████| 13/13 [00:13<00:00,  1.01s/it]
+Evaluated 400 examples
+[2026-07-21 07:50:57,712][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-21 07:50:57,713][evaluator][INFO] - Result for metric forget_truth_ratio:	0.4951789649259015
+[2026-07-21 07:50:57,721][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-21 07:50:57,722][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-21 07:50:57,738][metrics][INFO] - Evaluating forget_quality
+[2026-07-21 07:50:57,739][evaluator][INFO] - Result for metric forget_quality:	2.241003819205479e-17
+[2026-07-21 07:50:58,119][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.81it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.81it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.69it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.72it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.74it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.78it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.78it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.86it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.89it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.88it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.86it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.84it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  5.09it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.88it/s]
+Evaluated 400 examples
+[2026-07-21 07:51:00,785][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.744580078125
+[2026-07-21 07:51:01,061][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:06,  1.76it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:04,  2.21it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:01<00:04,  2.12it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:03,  2.37it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:02<00:03,  2.40it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:02<00:02,  2.41it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:03<00:02,  2.41it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:03<00:01,  2.54it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:03<00:01,  2.64it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:04<00:01,  2.23it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:04<00:00,  2.01it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:05<00:00,  2.01it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:05<00:00,  2.59it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:05<00:00,  2.34it/s]
+Evaluated 400 examples
+[2026-07-21 07:51:06,608][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.10923903386871697
+[2026-07-21 07:51:06,933][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.97it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  5.03it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:01,  5.04it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  5.06it/s]
Calculating loss:  38%|███▊      | 5/13 [00:00<00:01,  5.06it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.99it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.98it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.94it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.98it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.97it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.92it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.95it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  5.19it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  5.04it/s]
+Evaluated 400 examples
+[2026-07-21 07:51:09,781][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:07,  1.60it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:01<00:06,  1.72it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:01<00:05,  1.80it/s]
Calculating text similarity:  31%|███       | 4/13 [00:02<00:05,  1.71it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:02<00:04,  1.76it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:03<00:03,  1.78it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:04<00:03,  1.71it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:04<00:02,  1.69it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:05<00:02,  1.49it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:06<00:01,  1.59it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:06<00:01,  1.54it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:07<00:00,  1.62it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:07<00:00,  1.74it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:07<00:00,  1.68it/s]
+Evaluated 400 examples
+[2026-07-21 07:51:17,890][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.91it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.93it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.91it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.97it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.99it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.92it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.89it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.83it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.83it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.84it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.82it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.84it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  5.10it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.93it/s]
+Evaluated 400 examples
+[2026-07-21 07:51:20,800][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:11,  1.02it/s]
Calculating loss:  15%|█▌        | 2/13 [00:01<00:10,  1.03it/s]
Calculating loss:  23%|██▎       | 3/13 [00:02<00:09,  1.01it/s]
Calculating loss:  31%|███       | 4/13 [00:03<00:08,  1.02it/s]
Calculating loss:  38%|███▊      | 5/13 [00:04<00:07,  1.03it/s]
Calculating loss:  46%|████▌     | 6/13 [00:05<00:06,  1.02it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:06<00:05,  1.01it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:07<00:04,  1.00it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:08<00:04,  1.00s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:09<00:03,  1.00s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:10<00:02,  1.01s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:11<00:01,  1.01s/it]
Calculating loss: 100%|██████████| 13/13 [00:12<00:00,  1.04it/s]
Calculating loss: 100%|██████████| 13/13 [00:12<00:00,  1.02it/s]
+Evaluated 400 examples
+[2026-07-21 07:51:33,601][metrics][INFO] - Evaluating retain_Truth_Ratio
+[2026-07-21 07:51:33,969][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.41it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.84it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  7.99it/s]
+Evaluated 100 examples
+[2026-07-21 07:51:34,754][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.39it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.64it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.22it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.64it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.67it/s]
+Evaluated 100 examples
+[2026-07-21 07:51:36,251][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+[2026-07-21 07:51:36,518][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:01,  2.91it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  2.19it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:01<00:00,  2.40it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  2.76it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  2.62it/s]
+Evaluated 100 examples
+[2026-07-21 07:51:38,045][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-21 07:51:38,046][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-21 07:51:38,046][metrics][INFO] - Evaluating ra_Truth_Ratio
+[2026-07-21 07:51:38,313][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.40it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.89it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  8.66it/s]
+Evaluated 117 examples
+[2026-07-21 07:51:39,031][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.51it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.62it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.22it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.71it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.72it/s]
+Evaluated 117 examples
+[2026-07-21 07:51:40,502][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+[2026-07-21 07:51:40,771][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:01,  2.50it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  2.30it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:01<00:00,  2.27it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  2.28it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  2.30it/s]
+Evaluated 117 examples
+[2026-07-21 07:51:42,512][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-21 07:51:42,512][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-21 07:51:42,513][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-21 07:51:42,513][metrics][INFO] - Evaluating model_utility
+[2026-07-21 07:51:42,514][evaluator][INFO] - Result for metric model_utility:	0.5677726043908246
+[2026-07-21 07:51:43,074][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-21 07:51:43,089][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.28it/s]
 31%|███       | 4/13 [00:00<00:00, 11.48it/s]
 46%|████▌     | 6/13 [00:00<00:00, 11.80it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 11.93it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 12.19it/s]
 92%|█████████▏| 12/13 [00:01<00:00, 11.90it/s]
100%|██████████| 13/13 [00:01<00:00, 12.45it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 11.73it/s]
 31%|███       | 4/13 [00:00<00:00, 11.58it/s]
 46%|████▌     | 6/13 [00:00<00:00, 11.86it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 11.91it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 11.93it/s]
 92%|█████████▏| 12/13 [00:01<00:00, 12.04it/s]
100%|██████████| 13/13 [00:01<00:00, 12.34it/s]
+[2026-07-21 07:51:45,191][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-21 07:51:45,205][metrics][INFO] - Evaluating privleak
+[2026-07-21 07:51:45,205][evaluator][INFO] - Result for metric privleak:	-98.99475611362479
+[2026-07-21 07:51:45,496][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:00, 11.38it/s]
Calculating ES:  31%|███       | 4/13 [00:00<00:00, 10.07it/s]
Calculating ES:  46%|████▌     | 6/13 [00:00<00:00, 10.58it/s]
Calculating ES:  62%|██████▏   | 8/13 [00:00<00:00, 10.86it/s]
Calculating ES:  77%|███████▋  | 10/13 [00:00<00:00, 10.83it/s]
Calculating ES:  92%|█████████▏| 12/13 [00:01<00:00, 10.64it/s]
Calculating ES: 100%|██████████| 13/13 [00:01<00:00, 11.10it/s]
+Evaluated 400 examples
+[2026-07-21 07:51:46,668][evaluator][INFO] - Result for metric extraction_strength:	0.4922378536512631
diff --git a/logs/eval_IdkNLL_lr3e-05_alpha10_epoch5.log b/logs/eval_IdkNLL_lr3e-05_alpha10_epoch5.log
new file mode 100644
index 0000000000000000000000000000000000000000..c6ac31b462596c4940d61bff1ad4bdfbf8a7e63e
--- /dev/null
+++ b/logs/eval_IdkNLL_lr3e-05_alpha10_epoch5.log
@@ -0,0 +1,78 @@
+/opt/conda/envs/copygen/lib/python3.12/site-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.6.3) or chardet (6.0.0.post1)/charset_normalizer (3.4.4) doesn't match a supported version!
+  warnings.warn(
+[2026-07-22 17:16:24,031][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/IdkNLL_lr3e-05_alpha10_epoch5
+[2026-07-22 17:16:24,032][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-22 17:16:24,032][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/IdkNLL_lr3e-05_alpha10_epoch5/TOFU_EVAL.json
+[2026-07-22 17:16:24,032][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/IdkNLL_lr3e-05_alpha10_epoch5/TOFU_SUMMARY.json
+[2026-07-22 17:16:24,649][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:06,  1.82it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:03,  2.78it/s]
Calculating loss:  23%|██▎       | 3/13 [00:01<00:03,  3.28it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.62it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:02,  3.87it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.03it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.15it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:02<00:01,  4.29it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.33it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.38it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.37it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:03<00:00,  4.37it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.57it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.00it/s]
+Evaluated 400 examples
+[2026-07-22 17:16:28,617][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:13,  1.13s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:12,  1.12s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:11,  1.14s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:10,  1.14s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:09,  1.13s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.13s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.12s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.11s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:10<00:04,  1.11s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:11<00:03,  1.10s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:12<00:02,  1.11s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.11s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.06s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.10s/it]
+Evaluated 400 examples
+[2026-07-22 17:16:42,972][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-22 17:16:42,973][evaluator][INFO] - Result for metric forget_truth_ratio:	0.501259620393524
+[2026-07-22 17:16:42,977][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-22 17:16:42,977][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:16:42,984][metrics][INFO] - Evaluating forget_quality
+[2026-07-22 17:16:42,985][evaluator][INFO] - Result for metric forget_quality:	4.2392996435511593e-17
+[2026-07-22 17:16:43,732][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.35it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.42it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.29it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:02,  4.32it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.31it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.34it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.33it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.42it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.50it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.51it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.50it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.49it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.74it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.49it/s]
+Evaluated 400 examples
+[2026-07-22 17:16:46,629][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.7376806640625
+[2026-07-22 17:16:46,885][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:04,  2.79it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:03,  2.80it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:01<00:03,  3.08it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:02,  3.34it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:02,  3.36it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:01<00:01,  3.59it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:01<00:01,  3.96it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  4.07it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:02<00:00,  4.18it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:02<00:00,  3.79it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:03<00:00,  3.52it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:03<00:00,  3.43it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.73it/s]
+Evaluated 400 examples
+[2026-07-22 17:16:50,367][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.14538918415014265
+[2026-07-22 17:16:50,699][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.63it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.69it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.70it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.71it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.72it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.66it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.64it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.61it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.65it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.60it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.52it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.55it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.76it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.67it/s]
+Evaluated 400 examples
+[2026-07-22 17:16:53,716][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:03,  3.18it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:02,  3.69it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:02,  3.45it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:02,  3.63it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:02,  3.81it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:01<00:01,  3.86it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:01<00:01,  3.68it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  3.55it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:02<00:01,  3.45it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:02<00:00,  3.39it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:03<00:00,  3.35it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:03<00:00,  3.38it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.45it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.51it/s]
+Evaluated 400 examples
+[2026-07-22 17:16:57,806][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.61it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.61it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.57it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.63it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.66it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.59it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.55it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.50it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.50it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.51it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.48it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.51it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.75it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.60it/s]
+Evaluated 400 examples
+[2026-07-22 17:17:01,300][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:12,  1.07s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:11,  1.06s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:10,  1.07s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:09,  1.06s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:08,  1.06s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.07s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.09s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.10s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:09<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:10<00:03,  1.09s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:11<00:02,  1.10s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.10s/it]
Calculating loss: 100%|██████████| 13/13 [00:13<00:00,  1.05s/it]
Calculating loss: 100%|██████████| 13/13 [00:13<00:00,  1.07s/it]
+Evaluated 400 examples
+[2026-07-22 17:17:15,261][metrics][INFO] - Evaluating retain_Truth_Ratio
+[2026-07-22 17:17:16,504][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.10it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.53it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  7.51it/s]
+Evaluated 100 examples
+[2026-07-22 17:17:17,279][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  5.97it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.43it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.07it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.46it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.49it/s]
+Evaluated 100 examples
+[2026-07-22 17:17:18,887][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+[2026-07-22 17:17:19,122][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  5.15it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  4.73it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  5.14it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  5.94it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  5.55it/s]
+Evaluated 100 examples
+[2026-07-22 17:17:19,843][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-22 17:17:19,844][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:17:19,844][metrics][INFO] - Evaluating ra_Truth_Ratio
+[2026-07-22 17:17:20,415][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  7.69it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.44it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  8.15it/s]
+Evaluated 117 examples
+[2026-07-22 17:17:21,185][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.89it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.49it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.09it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.56it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.58it/s]
+Evaluated 117 examples
+[2026-07-22 17:17:22,737][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+[2026-07-22 17:17:22,982][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  4.54it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  4.37it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  4.10it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  3.95it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  4.06it/s]
+Evaluated 117 examples
+[2026-07-22 17:17:23,967][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-22 17:17:23,967][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:17:23,967][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-22 17:17:23,967][metrics][INFO] - Evaluating model_utility
+[2026-07-22 17:17:23,968][evaluator][INFO] - Result for metric model_utility:	0.5610351502264007
+[2026-07-22 17:17:25,769][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:17:25,779][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.22it/s]
 31%|███       | 4/13 [00:00<00:00, 12.09it/s]
 46%|████▌     | 6/13 [00:00<00:00, 12.95it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.47it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.93it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.74it/s]
100%|██████████| 13/13 [00:00<00:00, 14.04it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.56it/s]
 31%|███       | 4/13 [00:00<00:00, 13.42it/s]
 46%|████▌     | 6/13 [00:00<00:00, 13.91it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.93it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 14.03it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 14.13it/s]
100%|██████████| 13/13 [00:00<00:00, 14.44it/s]
+[2026-07-22 17:17:27,607][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:17:27,614][metrics][INFO] - Evaluating privleak
+[2026-07-22 17:17:27,614][evaluator][INFO] - Result for metric privleak:	-99.02613834924968
+[2026-07-22 17:17:28,178][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:00, 12.43it/s]
Calculating ES:  31%|███       | 4/13 [00:00<00:00, 11.31it/s]
Calculating ES:  46%|████▌     | 6/13 [00:00<00:00, 12.09it/s]
Calculating ES:  62%|██████▏   | 8/13 [00:00<00:00, 12.57it/s]
Calculating ES:  77%|███████▋  | 10/13 [00:00<00:00, 12.75it/s]
Calculating ES:  92%|█████████▏| 12/13 [00:00<00:00, 12.54it/s]
Calculating ES: 100%|██████████| 13/13 [00:01<00:00, 12.92it/s]
+Evaluated 400 examples
+[2026-07-22 17:17:29,185][evaluator][INFO] - Result for metric extraction_strength:	0.46257153663969325
diff --git a/logs/eval_IdkNLL_lr3e-05_alpha1_epoch10.log b/logs/eval_IdkNLL_lr3e-05_alpha1_epoch10.log
new file mode 100644
index 0000000000000000000000000000000000000000..02f360d15ca7b9329a9b85e30f010827943c4a47
--- /dev/null
+++ b/logs/eval_IdkNLL_lr3e-05_alpha1_epoch10.log
@@ -0,0 +1,78 @@
+/opt/conda/envs/copygen/lib/python3.12/site-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.6.3) or chardet (6.0.0.post1)/charset_normalizer (3.4.4) doesn't match a supported version!
+  warnings.warn(
+[2026-07-22 17:14:51,160][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/IdkNLL_lr3e-05_alpha1_epoch10
+[2026-07-22 17:14:51,161][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-22 17:14:51,161][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/IdkNLL_lr3e-05_alpha1_epoch10/TOFU_EVAL.json
+[2026-07-22 17:14:51,161][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/IdkNLL_lr3e-05_alpha1_epoch10/TOFU_SUMMARY.json
+[2026-07-22 17:14:51,829][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:04,  2.43it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:03,  3.31it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  3.68it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.90it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.10it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.20it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.29it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.38it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.40it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.44it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.42it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.40it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.60it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.20it/s]
+Evaluated 400 examples
+[2026-07-22 17:14:55,377][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:13,  1.11s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:12,  1.12s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:11,  1.15s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:10,  1.15s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:09,  1.13s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.13s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.12s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.10s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:10<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:11<00:03,  1.11s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:12<00:02,  1.11s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.11s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.06s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.10s/it]
+Evaluated 400 examples
+[2026-07-22 17:15:09,730][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-22 17:15:09,731][evaluator][INFO] - Result for metric forget_truth_ratio:	0.4949914861684177
+[2026-07-22 17:15:09,735][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-22 17:15:09,735][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:15:09,742][metrics][INFO] - Evaluating forget_quality
+[2026-07-22 17:15:09,742][evaluator][INFO] - Result for metric forget_quality:	3.2024272501633618e-18
+[2026-07-22 17:15:10,101][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.50it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.36it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.26it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:02,  4.32it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.36it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.41it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.40it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.50it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.54it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.54it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.51it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.49it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.74it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.51it/s]
+Evaluated 400 examples
+[2026-07-22 17:15:12,984][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.67515380859375
+[2026-07-22 17:15:13,235][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:01,  7.29it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:01,  9.28it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:00<00:00,  9.92it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:00<00:00, 10.41it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:00<00:00, 10.70it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:01<00:00,  8.68it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:01<00:00,  9.38it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:01<00:00,  9.50it/s]
+Evaluated 400 examples
+[2026-07-22 17:15:14,604][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.009139949321256842
+[2026-07-22 17:15:14,906][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.69it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.75it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.76it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.79it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.79it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.72it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.70it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.66it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.70it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.64it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.60it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.65it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.88it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.74it/s]
+Evaluated 400 examples
+[2026-07-22 17:15:17,887][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:03,  3.72it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:03,  3.52it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:03,  3.19it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:02,  3.16it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:02,  3.51it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:01<00:02,  3.48it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:02<00:01,  3.39it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  3.35it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:02<00:01,  3.37it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:02<00:00,  3.26it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:03<00:00,  3.05it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:03<00:00,  3.15it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.34it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.32it/s]
+Evaluated 400 examples
+[2026-07-22 17:15:22,075][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.48it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.49it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.47it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.55it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.61it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.56it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.54it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.50it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.51it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.53it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.51it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.54it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.78it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.59it/s]
+Evaluated 400 examples
+[2026-07-22 17:15:25,150][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:12,  1.06s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:11,  1.06s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:10,  1.07s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:09,  1.06s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:08,  1.06s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.08s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.09s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.09s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:09<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:10<00:03,  1.09s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:11<00:02,  1.10s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.10s/it]
Calculating loss: 100%|██████████| 13/13 [00:13<00:00,  1.05s/it]
Calculating loss: 100%|██████████| 13/13 [00:13<00:00,  1.08s/it]
+Evaluated 400 examples
+[2026-07-22 17:15:39,148][metrics][INFO] - Evaluating retain_Truth_Ratio
+[2026-07-22 17:15:39,488][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.13it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.57it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  7.57it/s]
+Evaluated 100 examples
+[2026-07-22 17:15:40,274][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.43it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.51it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.11it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.51it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.54it/s]
+Evaluated 100 examples
+[2026-07-22 17:15:41,846][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+[2026-07-22 17:15:42,152][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  5.32it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  4.54it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  5.13it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  6.17it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  5.65it/s]
+Evaluated 100 examples
+[2026-07-22 17:15:42,860][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-22 17:15:42,860][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:15:42,860][metrics][INFO] - Evaluating ra_Truth_Ratio
+[2026-07-22 17:15:43,110][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.01it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.51it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  8.25it/s]
+Evaluated 117 examples
+[2026-07-22 17:15:43,857][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.17it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.46it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.09it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.56it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.56it/s]
+Evaluated 117 examples
+[2026-07-22 17:15:45,418][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+[2026-07-22 17:15:45,664][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  3.75it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  3.77it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  4.13it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  4.13it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  4.05it/s]
+Evaluated 117 examples
+[2026-07-22 17:15:46,652][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-22 17:15:46,653][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:15:46,653][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-22 17:15:46,653][metrics][INFO] - Evaluating model_utility
+[2026-07-22 17:15:46,653][evaluator][INFO] - Result for metric model_utility:	0.4917842779052955
+[2026-07-22 17:15:47,155][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:15:47,166][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.23it/s]
 31%|███       | 4/13 [00:00<00:00, 11.88it/s]
 46%|████▌     | 6/13 [00:00<00:00, 12.55it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 12.97it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.41it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.21it/s]
100%|██████████| 13/13 [00:00<00:00, 13.58it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.05it/s]
 31%|███       | 4/13 [00:00<00:00, 12.92it/s]
 46%|████▌     | 6/13 [00:00<00:00, 13.35it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.35it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.40it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.50it/s]
100%|██████████| 13/13 [00:00<00:00, 13.82it/s]
+[2026-07-22 17:15:49,066][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:15:49,073][metrics][INFO] - Evaluating privleak
+[2026-07-22 17:15:49,073][evaluator][INFO] - Result for metric privleak:	-98.5989350126786
+[2026-07-22 17:15:49,326][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:00, 12.25it/s]
Calculating ES:  31%|███       | 4/13 [00:00<00:00, 11.05it/s]
Calculating ES:  46%|████▌     | 6/13 [00:00<00:00, 11.80it/s]
Calculating ES:  62%|██████▏   | 8/13 [00:00<00:00, 12.19it/s]
Calculating ES:  77%|███████▋  | 10/13 [00:00<00:00, 12.33it/s]
Calculating ES:  92%|█████████▏| 12/13 [00:01<00:00, 12.04it/s]
Calculating ES: 100%|██████████| 13/13 [00:01<00:00, 12.49it/s]
+Evaluated 400 examples
+[2026-07-22 17:15:50,368][evaluator][INFO] - Result for metric extraction_strength:	0.41322452248359304
diff --git a/logs/eval_IdkNLL_lr3e-05_alpha1_epoch5.log b/logs/eval_IdkNLL_lr3e-05_alpha1_epoch5.log
new file mode 100644
index 0000000000000000000000000000000000000000..12e47346e1cae88fee42dbafc3823aaac74760df
--- /dev/null
+++ b/logs/eval_IdkNLL_lr3e-05_alpha1_epoch5.log
@@ -0,0 +1,78 @@
+/opt/conda/envs/copygen/lib/python3.12/site-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.6.3) or chardet (6.0.0.post1)/charset_normalizer (3.4.4) doesn't match a supported version!
+  warnings.warn(
+[2026-07-22 17:08:22,993][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/IdkNLL_lr3e-05_alpha1_epoch5
+[2026-07-22 17:08:22,993][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-22 17:08:22,993][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/IdkNLL_lr3e-05_alpha1_epoch5/TOFU_EVAL.json
+[2026-07-22 17:08:22,994][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/IdkNLL_lr3e-05_alpha1_epoch5/TOFU_SUMMARY.json
+[2026-07-22 17:08:23,802][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:05,  2.27it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:03,  3.17it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  3.50it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.74it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:02,  3.94it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.06it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.18it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:02<00:01,  4.32it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.38it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.42it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.40it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.39it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.60it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.12it/s]
+Evaluated 400 examples
+[2026-07-22 17:08:27,465][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:13,  1.11s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:12,  1.12s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:11,  1.15s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:10,  1.15s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:09,  1.14s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.14s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.13s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:09<00:05,  1.11s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:10<00:04,  1.11s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:11<00:03,  1.11s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:12<00:02,  1.12s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.12s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.07s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.11s/it]
+Evaluated 400 examples
+[2026-07-22 17:08:41,905][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-22 17:08:41,905][evaluator][INFO] - Result for metric forget_truth_ratio:	0.4978633349891568
+[2026-07-22 17:08:41,909][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-22 17:08:41,910][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:08:41,916][metrics][INFO] - Evaluating forget_quality
+[2026-07-22 17:08:41,917][evaluator][INFO] - Result for metric forget_quality:	8.513756583610105e-19
+[2026-07-22 17:08:42,259][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.50it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.50it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.35it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:02,  4.38it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.41it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.46it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.46it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.55it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.59it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.58it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.55it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.52it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.76it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.56it/s]
+Evaluated 400 examples
+[2026-07-22 17:08:45,113][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.67819091796875
+[2026-07-22 17:08:45,420][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:01,  8.71it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:01,  9.66it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:00<00:00, 10.15it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:00<00:00, 10.61it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:00<00:00, 10.83it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:01<00:00,  9.24it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:01<00:00,  9.99it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:01<00:00, 10.00it/s]
+Evaluated 400 examples
+[2026-07-22 17:08:46,720][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.008150290861223104
+[2026-07-22 17:08:47,190][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.61it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.67it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.69it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.70it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.71it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.64it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.62it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.59it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.62it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.58it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.50it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.53it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.74it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.64it/s]
+Evaluated 400 examples
+[2026-07-22 17:08:50,251][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:03,  3.86it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:02,  3.88it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:02,  3.73it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:02,  3.83it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:01,  4.10it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:01<00:01,  4.18it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:01<00:01,  3.85it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  3.74it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:02<00:01,  3.70it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:02<00:00,  3.39it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:03<00:00,  3.30it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:03<00:00,  3.43it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.67it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.70it/s]
+Evaluated 400 examples
+[2026-07-22 17:08:54,012][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.59it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.59it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.56it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.62it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.65it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.57it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.53it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.48it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.48it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.49it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.46it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.48it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.71it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.58it/s]
+Evaluated 400 examples
+[2026-07-22 17:08:57,087][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:13,  1.09s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:11,  1.07s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:10,  1.08s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:09,  1.07s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:08,  1.06s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.08s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.09s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.10s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:09<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:10<00:03,  1.10s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:11<00:02,  1.10s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.10s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.05s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.08s/it]
+Evaluated 400 examples
+[2026-07-22 17:09:11,108][metrics][INFO] - Evaluating retain_Truth_Ratio
+[2026-07-22 17:09:11,462][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.16it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.56it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  7.54it/s]
+Evaluated 100 examples
+[2026-07-22 17:09:12,257][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.46it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.50it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.09it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.45it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.50it/s]
+Evaluated 100 examples
+[2026-07-22 17:09:13,854][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+[2026-07-22 17:09:14,107][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  5.98it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  4.93it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  5.53it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  6.18it/s]
+Evaluated 100 examples
+[2026-07-22 17:09:14,755][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-22 17:09:14,755][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:09:14,755][metrics][INFO] - Evaluating ra_Truth_Ratio
+[2026-07-22 17:09:14,992][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.08it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.60it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  8.38it/s]
+Evaluated 117 examples
+[2026-07-22 17:09:15,708][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.85it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.51it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.11it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.58it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.60it/s]
+Evaluated 117 examples
+[2026-07-22 17:09:17,250][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+[2026-07-22 17:09:17,546][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  4.18it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  4.96it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  5.21it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  5.03it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  4.97it/s]
+Evaluated 117 examples
+[2026-07-22 17:09:18,351][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-22 17:09:18,352][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:09:18,352][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-22 17:09:18,352][metrics][INFO] - Evaluating model_utility
+[2026-07-22 17:09:18,352][evaluator][INFO] - Result for metric model_utility:	0.44566065985412656
+[2026-07-22 17:09:18,942][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:09:18,949][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.04it/s]
 31%|███       | 4/13 [00:00<00:00, 12.24it/s]
 46%|████▌     | 6/13 [00:00<00:00, 12.90it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.21it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.62it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.36it/s]
100%|██████████| 13/13 [00:00<00:00, 13.77it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.18it/s]
 31%|███       | 4/13 [00:00<00:00, 12.95it/s]
 46%|████▌     | 6/13 [00:00<00:00, 13.34it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.35it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.35it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.44it/s]
100%|██████████| 13/13 [00:00<00:00, 13.78it/s]
+[2026-07-22 17:09:20,839][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:09:20,846][metrics][INFO] - Evaluating privleak
+[2026-07-22 17:09:20,846][evaluator][INFO] - Result for metric privleak:	-98.70016803082339
+[2026-07-22 17:09:21,132][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:00, 11.59it/s]
Calculating ES:  31%|███       | 4/13 [00:00<00:00, 10.44it/s]
Calculating ES:  46%|████▌     | 6/13 [00:00<00:00, 11.32it/s]
Calculating ES:  62%|██████▏   | 8/13 [00:00<00:00, 11.45it/s]
Calculating ES:  77%|███████▋  | 10/13 [00:00<00:00, 11.81it/s]
Calculating ES:  92%|█████████▏| 12/13 [00:01<00:00, 11.21it/s]
Calculating ES: 100%|██████████| 13/13 [00:01<00:00, 11.65it/s]
+Evaluated 400 examples
+[2026-07-22 17:09:22,248][evaluator][INFO] - Result for metric extraction_strength:	0.4052569846595127
diff --git a/logs/eval_IdkNLL_lr3e-05_alpha2_epoch10.log b/logs/eval_IdkNLL_lr3e-05_alpha2_epoch10.log
new file mode 100644
index 0000000000000000000000000000000000000000..839ad1182ef3ed7f3beefd814b4d9dcccf869bbb
--- /dev/null
+++ b/logs/eval_IdkNLL_lr3e-05_alpha2_epoch10.log
@@ -0,0 +1,78 @@
+/opt/conda/envs/copygen/lib/python3.12/site-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.6.3) or chardet (6.0.0.post1)/charset_normalizer (3.4.4) doesn't match a supported version!
+  warnings.warn(
+[2026-07-20 14:54:04,206][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/IdkNLL_lr3e-05_alpha2_epoch10
+[2026-07-20 14:54:04,207][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-20 14:54:04,207][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/IdkNLL_lr3e-05_alpha2_epoch10/TOFU_EVAL.json
+[2026-07-20 14:54:04,207][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/IdkNLL_lr3e-05_alpha2_epoch10/TOFU_SUMMARY.json
+[2026-07-20 14:54:04,932][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:09,  1.27it/s]
Calculating loss:  15%|█▌        | 2/13 [00:01<00:05,  2.09it/s]
Calculating loss:  23%|██▎       | 3/13 [00:01<00:03,  2.54it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:03,  2.79it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:02,  2.90it/s]
Calculating loss:  46%|████▌     | 6/13 [00:02<00:02,  3.07it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:02<00:01,  3.47it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:02<00:01,  3.83it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:03<00:01,  3.41it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:03<00:00,  3.11it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:03<00:00,  2.91it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:04<00:00,  2.90it/s]
Calculating loss: 100%|██████████| 13/13 [00:04<00:00,  2.83it/s]
Calculating loss: 100%|██████████| 13/13 [00:04<00:00,  2.87it/s]
+Evaluated 400 examples
+[2026-07-20 14:54:09,818][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:23,  1.99s/it]
Calculating loss:  15%|█▌        | 2/13 [00:03<00:18,  1.71s/it]
Calculating loss:  23%|██▎       | 3/13 [00:05<00:16,  1.63s/it]
Calculating loss:  31%|███       | 4/13 [00:06<00:14,  1.63s/it]
Calculating loss:  38%|███▊      | 5/13 [00:08<00:12,  1.57s/it]
Calculating loss:  46%|████▌     | 6/13 [00:09<00:10,  1.49s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:11<00:09,  1.51s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:12<00:07,  1.47s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:13<00:05,  1.44s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:15<00:04,  1.42s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:16<00:02,  1.40s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:18<00:01,  1.46s/it]
Calculating loss: 100%|██████████| 13/13 [00:19<00:00,  1.33s/it]
Calculating loss: 100%|██████████| 13/13 [00:19<00:00,  1.47s/it]
+Evaluated 400 examples
+[2026-07-20 14:54:28,938][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-20 14:54:28,939][evaluator][INFO] - Result for metric forget_truth_ratio:	0.4993750635808588
+[2026-07-20 14:54:28,947][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-20 14:54:28,948][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-20 14:54:28,964][metrics][INFO] - Evaluating forget_quality
+[2026-07-20 14:54:28,966][evaluator][INFO] - Result for metric forget_quality:	3.2024272501633618e-18
+[2026-07-20 14:54:29,472][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:04,  2.44it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:04,  2.53it/s]
Calculating loss:  23%|██▎       | 3/13 [00:01<00:04,  2.43it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:03,  2.49it/s]
Calculating loss:  38%|███▊      | 5/13 [00:02<00:03,  2.50it/s]
Calculating loss:  46%|████▌     | 6/13 [00:02<00:02,  2.52it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:02<00:02,  2.51it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:03<00:01,  2.56it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:03<00:01,  2.58it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:03<00:01,  2.66it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:04<00:00,  2.70it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:04<00:00,  3.11it/s]
Calculating loss: 100%|██████████| 13/13 [00:04<00:00,  3.62it/s]
Calculating loss: 100%|██████████| 13/13 [00:04<00:00,  2.80it/s]
+Evaluated 400 examples
+[2026-07-20 14:54:34,117][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.6839794921875
+[2026-07-20 14:54:34,550][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:04,  2.64it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:03,  3.00it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:02,  3.34it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:02,  3.34it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:02<00:04,  1.79it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:02<00:03,  2.10it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:02<00:02,  2.42it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:03<00:01,  2.63it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:03<00:01,  2.24it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:04<00:01,  1.80it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:04<00:00,  2.21it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:05<00:00,  2.43it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:05<00:00,  2.78it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:05<00:00,  2.44it/s]
+Evaluated 400 examples
+[2026-07-20 14:54:39,888][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.01505161696737646
+[2026-07-20 14:54:40,288][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:03,  3.36it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:03,  2.88it/s]
Calculating loss:  23%|██▎       | 3/13 [00:01<00:03,  2.68it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:03,  2.63it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:02,  2.72it/s]
Calculating loss:  46%|████▌     | 6/13 [00:02<00:02,  2.74it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:02<00:02,  2.61it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:02<00:01,  2.58it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:03<00:01,  2.40it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:03<00:01,  2.41it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:04<00:00,  2.60it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:04<00:00,  2.61it/s]
Calculating loss: 100%|██████████| 13/13 [00:04<00:00,  2.68it/s]
Calculating loss: 100%|██████████| 13/13 [00:04<00:00,  2.64it/s]
+Evaluated 400 examples
+[2026-07-20 14:54:45,615][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:01<00:13,  1.09s/it]
Calculating text similarity:  15%|█▌        | 2/13 [00:02<00:11,  1.04s/it]
Calculating text similarity:  23%|██▎       | 3/13 [00:03<00:11,  1.18s/it]
Calculating text similarity:  31%|███       | 4/13 [00:04<00:09,  1.06s/it]
Calculating text similarity:  38%|███▊      | 5/13 [00:05<00:08,  1.11s/it]
Calculating text similarity:  46%|████▌     | 6/13 [00:06<00:08,  1.16s/it]
Calculating text similarity:  54%|█████▍    | 7/13 [00:08<00:07,  1.24s/it]
Calculating text similarity:  62%|██████▏   | 8/13 [00:09<00:05,  1.17s/it]
Calculating text similarity:  69%|██████▉   | 9/13 [00:10<00:05,  1.32s/it]
Calculating text similarity:  77%|███████▋  | 10/13 [00:12<00:03,  1.27s/it]
Calculating text similarity:  85%|████████▍ | 11/13 [00:12<00:02,  1.17s/it]
Calculating text similarity:  92%|█████████▏| 12/13 [00:13<00:01,  1.03s/it]
Calculating text similarity: 100%|██████████| 13/13 [00:14<00:00,  1.11it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:14<00:00,  1.10s/it]
+Evaluated 400 examples
+[2026-07-20 14:55:00,609][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:03,  3.36it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:03,  3.19it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:03,  3.23it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.29it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:02,  3.30it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:02,  3.30it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:02<00:01,  3.24it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:02<00:01,  3.16it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:01,  3.19it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:03<00:00,  3.18it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:03<00:00,  3.12it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:03<00:00,  3.26it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  3.48it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  3.29it/s]
+Evaluated 400 examples
+[2026-07-20 14:55:05,074][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:17,  1.47s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:15,  1.41s/it]
Calculating loss:  23%|██▎       | 3/13 [00:04<00:14,  1.44s/it]
Calculating loss:  31%|███       | 4/13 [00:05<00:12,  1.42s/it]
Calculating loss:  38%|███▊      | 5/13 [00:07<00:11,  1.42s/it]
Calculating loss:  46%|████▌     | 6/13 [00:08<00:10,  1.44s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:10<00:09,  1.52s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:12<00:08,  1.61s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:13<00:05,  1.49s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:14<00:04,  1.47s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:16<00:03,  1.51s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:17<00:01,  1.55s/it]
Calculating loss: 100%|██████████| 13/13 [00:19<00:00,  1.45s/it]
Calculating loss: 100%|██████████| 13/13 [00:19<00:00,  1.48s/it]
+Evaluated 400 examples
+[2026-07-20 14:55:24,256][metrics][INFO] - Evaluating retain_Truth_Ratio
+[2026-07-20 14:55:24,881][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  7.65it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  4.97it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  4.76it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  5.05it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  5.11it/s]
+Evaluated 100 examples
+[2026-07-20 14:55:26,179][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  4.25it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:01,  1.86it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  1.56it/s]
Calculating loss: 100%|██████████| 4/4 [00:02<00:00,  2.03it/s]
Calculating loss: 100%|██████████| 4/4 [00:02<00:00,  1.98it/s]
+Evaluated 100 examples
+[2026-07-20 14:55:28,198][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+[2026-07-20 14:55:28,569][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:01,  1.76it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:01<00:01,  1.73it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:01<00:00,  1.63it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:02<00:00,  1.61it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:02<00:00,  1.64it/s]
+Evaluated 100 examples
+[2026-07-20 14:55:31,010][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-20 14:55:31,010][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-20 14:55:31,010][metrics][INFO] - Evaluating ra_Truth_Ratio
+[2026-07-20 14:55:31,452][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  8.49it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  4.83it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  4.43it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  5.75it/s]
+Evaluated 117 examples
+[2026-07-20 14:55:32,687][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  3.58it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.02it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  1.81it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.15it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.13it/s]
+Evaluated 117 examples
+[2026-07-20 14:55:34,566][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+[2026-07-20 14:55:34,956][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:01,  2.11it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:01<00:01,  1.25it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:02<00:00,  1.30it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:03<00:00,  1.15it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:03<00:00,  1.23it/s]
+Evaluated 117 examples
+[2026-07-20 14:55:38,219][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-20 14:55:38,220][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-20 14:55:38,220][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-20 14:55:38,220][metrics][INFO] - Evaluating model_utility
+[2026-07-20 14:55:38,221][evaluator][INFO] - Result for metric model_utility:	0.5450860673113619
+[2026-07-20 14:55:39,086][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-20 14:55:39,175][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
  8%|▊         | 1/13 [00:00<00:01,  6.66it/s]
 15%|█▌        | 2/13 [00:00<00:01,  7.82it/s]
 31%|███       | 4/13 [00:00<00:00,  9.40it/s]
 46%|████▌     | 6/13 [00:00<00:00, 10.47it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 10.98it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 11.46it/s]
 92%|█████████▏| 12/13 [00:01<00:00, 11.35it/s]
100%|██████████| 13/13 [00:01<00:00, 11.13it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 11.43it/s]
 31%|███       | 4/13 [00:00<00:00, 11.27it/s]
 46%|████▌     | 6/13 [00:00<00:00, 11.63it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 11.69it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 11.72it/s]
 92%|█████████▏| 12/13 [00:01<00:00,  6.36it/s]
100%|██████████| 13/13 [00:01<00:00,  6.37it/s]
100%|██████████| 13/13 [00:01<00:00,  8.08it/s]
+[2026-07-20 14:55:41,960][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-20 14:55:41,975][metrics][INFO] - Evaluating privleak
+[2026-07-20 14:55:41,975][evaluator][INFO] - Result for metric privleak:	-98.66068715374693
+[2026-07-20 14:55:42,404][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:   8%|▊         | 1/13 [00:00<00:05,  2.24it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:04,  2.65it/s]
Calculating ES:  23%|██▎       | 3/13 [00:01<00:04,  2.28it/s]
Calculating ES:  31%|███       | 4/13 [00:01<00:03,  2.36it/s]
Calculating ES:  38%|███▊      | 5/13 [00:01<00:02,  2.71it/s]
Calculating ES:  46%|████▌     | 6/13 [00:02<00:02,  2.56it/s]
Calculating ES:  54%|█████▍    | 7/13 [00:02<00:01,  3.08it/s]
Calculating ES:  62%|██████▏   | 8/13 [00:02<00:01,  3.95it/s]
Calculating ES:  69%|██████▉   | 9/13 [00:02<00:00,  4.23it/s]
Calculating ES:  77%|███████▋  | 10/13 [00:03<00:00,  4.44it/s]
Calculating ES:  85%|████████▍ | 11/13 [00:04<00:01,  1.91it/s]
Calculating ES:  92%|█████████▏| 12/13 [00:04<00:00,  1.74it/s]
Calculating ES: 100%|██████████| 13/13 [00:05<00:00,  1.84it/s]
Calculating ES: 100%|██████████| 13/13 [00:05<00:00,  2.39it/s]
+Evaluated 400 examples
+[2026-07-20 14:55:47,850][evaluator][INFO] - Result for metric extraction_strength:	0.4109878851926348
diff --git a/logs/eval_IdkNLL_lr3e-05_alpha2_epoch5.log b/logs/eval_IdkNLL_lr3e-05_alpha2_epoch5.log
new file mode 100644
index 0000000000000000000000000000000000000000..8104229cef9c6d5cab63494c1688d3d85b68a6fb
--- /dev/null
+++ b/logs/eval_IdkNLL_lr3e-05_alpha2_epoch5.log
@@ -0,0 +1,78 @@
+/opt/conda/envs/copygen/lib/python3.12/site-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.6.3) or chardet (6.0.0.post1)/charset_normalizer (3.4.4) doesn't match a supported version!
+  warnings.warn(
+[2026-07-22 17:17:54,650][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/IdkNLL_lr3e-05_alpha2_epoch5
+[2026-07-22 17:17:54,651][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-22 17:17:54,651][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/IdkNLL_lr3e-05_alpha2_epoch5/TOFU_EVAL.json
+[2026-07-22 17:17:54,651][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/IdkNLL_lr3e-05_alpha2_epoch5/TOFU_SUMMARY.json
+[2026-07-22 17:18:00,407][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:04,  2.47it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:03,  3.34it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  3.69it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.91it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.10it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.19it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.28it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.39it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.43it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.46it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.44it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.42it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.62it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.22it/s]
+Evaluated 400 examples
+[2026-07-22 17:18:04,333][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:13,  1.12s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:12,  1.12s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:11,  1.16s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:10,  1.15s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:09,  1.14s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.13s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.12s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.11s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:10<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:11<00:03,  1.10s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:12<00:02,  1.12s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.12s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.06s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.11s/it]
+Evaluated 400 examples
+[2026-07-22 17:18:18,714][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-22 17:18:18,715][evaluator][INFO] - Result for metric forget_truth_ratio:	0.5028844227401504
+[2026-07-22 17:18:18,719][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-22 17:18:18,719][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:18:18,726][metrics][INFO] - Evaluating forget_quality
+[2026-07-22 17:18:18,726][evaluator][INFO] - Result for metric forget_quality:	6.159412995752159e-18
+[2026-07-22 17:18:19,097][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.51it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.51it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.35it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:02,  4.34it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.38it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.43it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.43it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.52it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.56it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.55it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.52it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.50it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.74it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.53it/s]
+Evaluated 400 examples
+[2026-07-22 17:18:21,965][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.68916015625
+[2026-07-22 17:18:22,232][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:02,  5.52it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:01,  6.36it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:01,  7.63it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:00<00:01,  6.07it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:00<00:00,  7.48it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:01<00:00,  7.67it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:01<00:00,  7.22it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:01<00:00,  7.38it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:01<00:00,  7.47it/s]
+Evaluated 400 examples
+[2026-07-22 17:18:23,973][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.016458309696891105
+[2026-07-22 17:18:24,251][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.67it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.72it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.74it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.77it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.77it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.70it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.69it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.65it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.69it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.67it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.62it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.66it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.88it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.74it/s]
+Evaluated 400 examples
+[2026-07-22 17:18:27,246][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:03,  3.72it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:03,  3.47it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:02,  3.34it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:02,  3.63it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:02,  3.92it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:01<00:01,  3.74it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:01<00:01,  3.59it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  3.66it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:02<00:01,  3.62it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:02<00:00,  3.24it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:03<00:00,  3.05it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:03<00:00,  2.98it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.21it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.38it/s]
+Evaluated 400 examples
+[2026-07-22 17:18:31,343][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.57it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.48it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.45it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.52it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.57it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.53it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.52it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.48it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.49it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.52it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.50it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.53it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.77it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.57it/s]
+Evaluated 400 examples
+[2026-07-22 17:18:34,545][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:12,  1.06s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:11,  1.06s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:10,  1.07s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:09,  1.07s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:08,  1.06s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.08s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.09s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.09s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:09<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:10<00:03,  1.10s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:11<00:02,  1.10s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.10s/it]
Calculating loss: 100%|██████████| 13/13 [00:13<00:00,  1.05s/it]
Calculating loss: 100%|██████████| 13/13 [00:13<00:00,  1.08s/it]
+Evaluated 400 examples
+[2026-07-22 17:18:48,545][metrics][INFO] - Evaluating retain_Truth_Ratio
+[2026-07-22 17:18:48,957][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.06it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.51it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  7.49it/s]
+Evaluated 100 examples
+[2026-07-22 17:18:49,846][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.52it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.51it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.09it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.49it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.53it/s]
+Evaluated 100 examples
+[2026-07-22 17:18:51,429][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+[2026-07-22 17:18:51,732][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  4.66it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  5.23it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  5.50it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  6.33it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  5.87it/s]
+Evaluated 100 examples
+[2026-07-22 17:18:52,414][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-22 17:18:52,414][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:18:52,414][metrics][INFO] - Evaluating ra_Truth_Ratio
+[2026-07-22 17:18:52,709][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.11it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.53it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  8.29it/s]
+Evaluated 117 examples
+[2026-07-22 17:18:53,480][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.26it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.44it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.07it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.53it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.54it/s]
+Evaluated 117 examples
+[2026-07-22 17:18:55,055][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+[2026-07-22 17:18:55,580][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  3.45it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  3.27it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  3.43it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  3.54it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  3.48it/s]
+Evaluated 117 examples
+[2026-07-22 17:18:56,730][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-22 17:18:56,730][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:18:56,730][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-22 17:18:56,730][metrics][INFO] - Evaluating model_utility
+[2026-07-22 17:18:56,731][evaluator][INFO] - Result for metric model_utility:	0.5224317236980688
+[2026-07-22 17:18:57,399][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:18:57,406][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.68it/s]
 31%|███       | 4/13 [00:00<00:00, 12.53it/s]
 46%|████▌     | 6/13 [00:00<00:00, 13.15it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.63it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 14.04it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.83it/s]
100%|██████████| 13/13 [00:00<00:00, 14.21it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.73it/s]
 31%|███       | 4/13 [00:00<00:00, 13.61it/s]
 46%|████▌     | 6/13 [00:00<00:00, 14.01it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 14.05it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 14.09it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 14.21it/s]
100%|██████████| 13/13 [00:00<00:00, 14.55it/s]
+[2026-07-22 17:18:59,217][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:18:59,223][metrics][INFO] - Evaluating privleak
+[2026-07-22 17:18:59,223][evaluator][INFO] - Result for metric privleak:	-98.77305580388764
+[2026-07-22 17:18:59,487][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:00, 12.35it/s]
Calculating ES:  31%|███       | 4/13 [00:00<00:00, 11.17it/s]
Calculating ES:  46%|████▌     | 6/13 [00:00<00:00, 11.94it/s]
Calculating ES:  62%|██████▏   | 8/13 [00:00<00:00, 12.33it/s]
Calculating ES:  77%|███████▋  | 10/13 [00:00<00:00, 12.51it/s]
Calculating ES:  92%|█████████▏| 12/13 [00:00<00:00, 12.33it/s]
Calculating ES: 100%|██████████| 13/13 [00:01<00:00, 12.71it/s]
+Evaluated 400 examples
+[2026-07-22 17:19:00,510][evaluator][INFO] - Result for metric extraction_strength:	0.4085078851067708
diff --git a/logs/eval_IdkNLL_lr3e-05_alpha5_epoch10.log b/logs/eval_IdkNLL_lr3e-05_alpha5_epoch10.log
new file mode 100644
index 0000000000000000000000000000000000000000..e789982a381c0c7021eb4fd97f9277a69c7019ce
--- /dev/null
+++ b/logs/eval_IdkNLL_lr3e-05_alpha5_epoch10.log
@@ -0,0 +1,78 @@
+/opt/conda/envs/copygen/lib/python3.12/site-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.6.3) or chardet (6.0.0.post1)/charset_normalizer (3.4.4) doesn't match a supported version!
+  warnings.warn(
+[2026-07-22 17:16:20,718][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/IdkNLL_lr3e-05_alpha5_epoch10
+[2026-07-22 17:16:20,719][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-22 17:16:20,719][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/IdkNLL_lr3e-05_alpha5_epoch10/TOFU_EVAL.json
+[2026-07-22 17:16:20,719][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/IdkNLL_lr3e-05_alpha5_epoch10/TOFU_SUMMARY.json
+[2026-07-22 17:16:21,393][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:04,  2.44it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:03,  3.32it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  3.67it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.90it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.09it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.19it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.29it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.41it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.44it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.47it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.45it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.42it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.63it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.22it/s]
+Evaluated 400 examples
+[2026-07-22 17:16:24,860][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:13,  1.12s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:12,  1.12s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:11,  1.14s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:10,  1.15s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:09,  1.14s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.13s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.12s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.10s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:10<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:11<00:03,  1.10s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:12<00:02,  1.12s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.12s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.06s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.10s/it]
+Evaluated 400 examples
+[2026-07-22 17:16:39,222][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-22 17:16:39,222][evaluator][INFO] - Result for metric forget_truth_ratio:	0.4968766294464798
+[2026-07-22 17:16:39,226][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-22 17:16:39,227][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:16:39,233][metrics][INFO] - Evaluating forget_quality
+[2026-07-22 17:16:39,234][evaluator][INFO] - Result for metric forget_quality:	6.159412995752159e-18
+[2026-07-22 17:16:39,578][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.52it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.51it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.36it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:02,  4.40it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.43it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.48it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.47it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.53it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.57it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.56it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.52it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.50it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.75it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.55it/s]
+Evaluated 400 examples
+[2026-07-22 17:16:42,434][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.71641357421875
+[2026-07-22 17:16:42,727][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:02,  5.05it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:01,  6.03it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:01,  7.05it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:00<00:01,  6.46it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:00<00:01,  6.08it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:01<00:00,  7.52it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:01<00:00,  6.69it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:01<00:00,  6.35it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:01<00:00,  5.90it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:02<00:00,  4.89it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:02<00:00,  6.12it/s]
+Evaluated 400 examples
+[2026-07-22 17:16:44,853][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.037324577993518855
+[2026-07-22 17:16:45,115][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.44it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.57it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.65it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.70it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.73it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.68it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.68it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.65it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.68it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.67it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.62it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.66it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.89it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.71it/s]
+Evaluated 400 examples
+[2026-07-22 17:16:53,393][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:04,  3.00it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:03,  3.48it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:02,  3.59it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:02,  3.68it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:02,  3.70it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:01<00:01,  3.57it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:02<00:01,  3.37it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  3.38it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:02<00:01,  3.48it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:02<00:00,  3.53it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:03<00:00,  3.27it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:03<00:00,  3.37it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.63it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.50it/s]
+Evaluated 400 examples
+[2026-07-22 17:16:58,424][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.61it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.62it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.57it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.63it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.67it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.59it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.56it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.50it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.50it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.52it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.45it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.45it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.67it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.58it/s]
+Evaluated 400 examples
+[2026-07-22 17:17:02,501][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:12,  1.05s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:11,  1.05s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:10,  1.07s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:09,  1.06s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:08,  1.06s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.07s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.09s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.09s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:09<00:04,  1.09s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:10<00:03,  1.09s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:11<00:02,  1.10s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.10s/it]
Calculating loss: 100%|██████████| 13/13 [00:13<00:00,  1.05s/it]
Calculating loss: 100%|██████████| 13/13 [00:13<00:00,  1.07s/it]
+Evaluated 400 examples
+[2026-07-22 17:17:16,442][metrics][INFO] - Evaluating retain_Truth_Ratio
+[2026-07-22 17:17:16,790][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  7.65it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.33it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  6.72it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  6.77it/s]
+Evaluated 100 examples
+[2026-07-22 17:17:17,631][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.14it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.47it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.09it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.51it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.53it/s]
+Evaluated 100 examples
+[2026-07-22 17:17:19,211][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+[2026-07-22 17:17:21,022][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  5.62it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  4.69it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  5.17it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  5.52it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  5.35it/s]
+Evaluated 100 examples
+[2026-07-22 17:17:21,771][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-22 17:17:21,771][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:17:21,771][metrics][INFO] - Evaluating ra_Truth_Ratio
+[2026-07-22 17:17:22,008][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.08it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.58it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  8.32it/s]
+Evaluated 117 examples
+[2026-07-22 17:17:24,084][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.42it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.46it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.08it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.52it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.54it/s]
+Evaluated 117 examples
+[2026-07-22 17:17:25,659][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+[2026-07-22 17:17:25,908][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  4.19it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  4.10it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  3.41it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  3.67it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  3.70it/s]
+Evaluated 117 examples
+[2026-07-22 17:17:26,989][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-22 17:17:26,989][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:17:26,989][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-22 17:17:26,989][metrics][INFO] - Evaluating model_utility
+[2026-07-22 17:17:26,989][evaluator][INFO] - Result for metric model_utility:	0.5687425441916057
+[2026-07-22 17:17:28,560][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:17:28,568][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.60it/s]
 31%|███       | 4/13 [00:00<00:00, 12.40it/s]
 46%|████▌     | 6/13 [00:00<00:00, 12.99it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.31it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.56it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.23it/s]
100%|██████████| 13/13 [00:00<00:00, 13.74it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.01it/s]
 31%|███       | 4/13 [00:00<00:00, 12.87it/s]
 46%|████▌     | 6/13 [00:00<00:00, 13.23it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.21it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.27it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.13it/s]
100%|██████████| 13/13 [00:00<00:00, 13.58it/s]
+[2026-07-22 17:17:30,475][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:17:30,482][metrics][INFO] - Evaluating privleak
+[2026-07-22 17:17:30,482][evaluator][INFO] - Result for metric privleak:	-98.86214085985509
+[2026-07-22 17:17:33,634][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:00, 11.44it/s]
Calculating ES:  31%|███       | 4/13 [00:00<00:00, 10.44it/s]
Calculating ES:  46%|████▌     | 6/13 [00:00<00:00, 11.19it/s]
Calculating ES:  62%|██████▏   | 8/13 [00:00<00:00, 11.65it/s]
Calculating ES:  77%|███████▋  | 10/13 [00:00<00:00, 11.83it/s]
Calculating ES:  92%|█████████▏| 12/13 [00:01<00:00, 11.67it/s]
Calculating ES: 100%|██████████| 13/13 [00:01<00:00, 12.00it/s]
+Evaluated 400 examples
+[2026-07-22 17:17:34,718][evaluator][INFO] - Result for metric extraction_strength:	0.4545462262677516
diff --git a/logs/eval_IdkNLL_lr3e-05_alpha5_epoch5.log b/logs/eval_IdkNLL_lr3e-05_alpha5_epoch5.log
new file mode 100644
index 0000000000000000000000000000000000000000..9f1cc3efb1e9cf52ee617b7bb3b6283de12013ad
--- /dev/null
+++ b/logs/eval_IdkNLL_lr3e-05_alpha5_epoch5.log
@@ -0,0 +1,78 @@
+/opt/conda/envs/copygen/lib/python3.12/site-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.6.3) or chardet (6.0.0.post1)/charset_normalizer (3.4.4) doesn't match a supported version!
+  warnings.warn(
+[2026-07-22 17:09:46,598][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/IdkNLL_lr3e-05_alpha5_epoch5
+[2026-07-22 17:09:46,599][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-22 17:09:46,599][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/IdkNLL_lr3e-05_alpha5_epoch5/TOFU_EVAL.json
+[2026-07-22 17:09:46,599][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/IdkNLL_lr3e-05_alpha5_epoch5/TOFU_SUMMARY.json
+[2026-07-22 17:09:52,693][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:06,  1.87it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:03,  2.82it/s]
Calculating loss:  23%|██▎       | 3/13 [00:01<00:03,  3.30it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.62it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:02,  3.88it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.02it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.15it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:02<00:01,  4.29it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.35it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.39it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.37it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:03<00:00,  4.36it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.56it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.01it/s]
+Evaluated 400 examples
+[2026-07-22 17:09:56,400][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:13,  1.15s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:12,  1.13s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:11,  1.15s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:10,  1.15s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:09,  1.14s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.13s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.13s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:09<00:05,  1.11s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:10<00:04,  1.11s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:11<00:03,  1.11s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:12<00:02,  1.12s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.12s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.06s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.11s/it]
+Evaluated 400 examples
+[2026-07-22 17:10:10,830][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-22 17:10:10,830][evaluator][INFO] - Result for metric forget_truth_ratio:	0.5024821186592119
+[2026-07-22 17:10:10,835][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-22 17:10:10,835][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:10:10,843][metrics][INFO] - Evaluating forget_quality
+[2026-07-22 17:10:10,844][evaluator][INFO] - Result for metric forget_quality:	6.159412995752159e-18
+[2026-07-22 17:10:11,265][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.46it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.45it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.30it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:02,  4.32it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.36it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.41it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.41it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.50it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.54it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.48it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.43it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.39it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.62it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.47it/s]
+Evaluated 400 examples
+[2026-07-22 17:10:14,174][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.7125732421875
+[2026-07-22 17:10:14,551][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:02,  4.35it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:01,  5.69it/s]
Calculating text similarity:  31%|███       | 4/13 [00:00<00:01,  6.77it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:00<00:01,  5.74it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:01<00:00,  7.40it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:01<00:00,  7.72it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:01<00:00,  7.28it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:01<00:00,  6.27it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:01<00:00,  5.94it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:01<00:00,  6.71it/s]
+Evaluated 400 examples
+[2026-07-22 17:10:16,490][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.03528257578687019
+[2026-07-22 17:10:16,741][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.67it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.73it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.74it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.76it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.76it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.69it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.64it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.60it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.64it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.63it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.57it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.61it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.83it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.70it/s]
+Evaluated 400 examples
+[2026-07-22 17:10:19,804][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:04,  2.93it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:03,  3.16it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:03,  3.02it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:02,  3.00it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:02,  3.19it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:01<00:02,  2.95it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:02<00:01,  3.01it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  3.08it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:02<00:01,  3.12it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:03<00:00,  3.20it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:03<00:00,  3.07it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:03<00:00,  3.11it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:04<00:00,  3.34it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:04<00:00,  3.14it/s]
+Evaluated 400 examples
+[2026-07-22 17:10:24,244][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.59it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.62it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.59it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.65it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.69it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.61it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.57it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.52it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.51it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.49it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.46it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.49it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.73it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.60it/s]
+Evaluated 400 examples
+[2026-07-22 17:10:27,324][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:12,  1.07s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:11,  1.07s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:10,  1.09s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:09,  1.07s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:08,  1.06s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.07s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.09s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.09s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:09<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:10<00:03,  1.10s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:12<00:02,  1.11s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.10s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.05s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.08s/it]
+Evaluated 400 examples
+[2026-07-22 17:10:41,352][metrics][INFO] - Evaluating retain_Truth_Ratio
+[2026-07-22 17:10:41,698][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.24it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.60it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  7.63it/s]
+Evaluated 100 examples
+[2026-07-22 17:10:42,458][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.38it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.45it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.07it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.43it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.48it/s]
+Evaluated 100 examples
+[2026-07-22 17:10:44,073][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+[2026-07-22 17:10:44,334][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  4.66it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  4.86it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  5.12it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  6.15it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  5.64it/s]
+Evaluated 100 examples
+[2026-07-22 17:10:45,045][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-22 17:10:45,045][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:10:45,045][metrics][INFO] - Evaluating ra_Truth_Ratio
+[2026-07-22 17:10:45,288][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.03it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.53it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  8.25it/s]
+Evaluated 117 examples
+[2026-07-22 17:10:46,045][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.46it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.42it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.05it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.53it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.53it/s]
+Evaluated 117 examples
+[2026-07-22 17:10:47,624][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+[2026-07-22 17:10:47,891][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  4.12it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  4.49it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  4.45it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  4.38it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  4.39it/s]
+Evaluated 117 examples
+[2026-07-22 17:10:48,803][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-22 17:10:48,804][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:10:48,804][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-22 17:10:48,804][metrics][INFO] - Evaluating model_utility
+[2026-07-22 17:10:48,804][evaluator][INFO] - Result for metric model_utility:	0.5514159224481929
+[2026-07-22 17:10:49,396][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:10:49,403][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.99it/s]
 31%|███       | 4/13 [00:00<00:00, 12.68it/s]
 46%|████▌     | 6/13 [00:00<00:00, 13.29it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.64it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.98it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.70it/s]
100%|██████████| 13/13 [00:00<00:00, 14.19it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.40it/s]
 31%|███       | 4/13 [00:00<00:00, 13.13it/s]
 46%|████▌     | 6/13 [00:00<00:00, 13.51it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.53it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.42it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 12.88it/s]
100%|██████████| 13/13 [00:00<00:00, 13.47it/s]
+[2026-07-22 17:10:51,287][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:10:51,294][metrics][INFO] - Evaluating privleak
+[2026-07-22 17:10:51,295][evaluator][INFO] - Result for metric privleak:	-98.89048610493563
+[2026-07-22 17:10:51,557][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:00, 11.72it/s]
Calculating ES:  31%|███       | 4/13 [00:00<00:00, 10.71it/s]
Calculating ES:  46%|████▌     | 6/13 [00:00<00:00, 11.48it/s]
Calculating ES:  62%|██████▏   | 8/13 [00:00<00:00, 11.98it/s]
Calculating ES:  77%|███████▋  | 10/13 [00:00<00:00, 11.81it/s]
Calculating ES:  92%|█████████▏| 12/13 [00:01<00:00, 11.05it/s]
Calculating ES: 100%|██████████| 13/13 [00:01<00:00, 11.75it/s]
+Evaluated 400 examples
+[2026-07-22 17:10:52,664][evaluator][INFO] - Result for metric extraction_strength:	0.43525861530382004
diff --git a/logs/eval_IdkNLL_lr4e-05_alpha10_epoch10.log b/logs/eval_IdkNLL_lr4e-05_alpha10_epoch10.log
new file mode 100644
index 0000000000000000000000000000000000000000..61cfbf0cbbca48470d1781eac53d92208a6ec1f0
--- /dev/null
+++ b/logs/eval_IdkNLL_lr4e-05_alpha10_epoch10.log
@@ -0,0 +1,76 @@
+[2026-07-21 07:50:36,858][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/IdkNLL_lr4e-05_alpha10_epoch10
+[2026-07-21 07:50:36,860][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-21 07:50:36,860][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/IdkNLL_lr4e-05_alpha10_epoch10/TOFU_EVAL.json
+[2026-07-21 07:50:36,860][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/IdkNLL_lr4e-05_alpha10_epoch10/TOFU_SUMMARY.json
+[2026-07-21 07:50:37,515][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:07,  1.51it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:04,  2.51it/s]
Calculating loss:  23%|██▎       | 3/13 [00:01<00:03,  3.16it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.61it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:02,  3.97it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.19it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.37it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:02<00:01,  4.53it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.62it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.68it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.68it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.68it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.83it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.09it/s]
+Evaluated 400 examples
+[2026-07-21 07:50:41,048][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:12,  1.03s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:11,  1.03s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:10,  1.05s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:09,  1.04s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:08,  1.04s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.03s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.03s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.01s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:09<00:04,  1.01s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:10<00:03,  1.01s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:11<00:02,  1.02s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:12<00:01,  1.02s/it]
Calculating loss: 100%|██████████| 13/13 [00:13<00:00,  1.03it/s]
Calculating loss: 100%|██████████| 13/13 [00:13<00:00,  1.01s/it]
+Evaluated 400 examples
+[2026-07-21 07:50:54,200][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-21 07:50:54,201][evaluator][INFO] - Result for metric forget_truth_ratio:	0.5113293245568581
+[2026-07-21 07:50:54,209][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-21 07:50:54,210][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-21 07:50:54,225][metrics][INFO] - Evaluating forget_quality
+[2026-07-21 07:50:54,227][evaluator][INFO] - Result for metric forget_quality:	1.1781277220965948e-17
+[2026-07-21 07:50:54,604][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.79it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.80it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.67it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.70it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.72it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.76it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.76it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.84it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.87it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.85it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.84it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.82it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  5.08it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.86it/s]
+Evaluated 400 examples
+[2026-07-21 07:50:57,279][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.6776708984375
+[2026-07-21 07:50:57,752][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:05,  2.33it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:04,  2.74it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:01<00:04,  2.18it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:03,  2.98it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:03,  2.57it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:02<00:02,  2.72it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:02<00:02,  2.79it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  2.83it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:03<00:01,  2.84it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:03<00:01,  2.89it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:04<00:00,  2.83it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:04<00:00,  2.54it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:04<00:00,  3.18it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:04<00:00,  2.81it/s]
+Evaluated 400 examples
+[2026-07-21 07:51:02,383][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.04984328424422209
+[2026-07-21 07:51:02,666][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.95it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  5.02it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:01,  5.03it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  5.05it/s]
Calculating loss:  38%|███▊      | 5/13 [00:00<00:01,  5.05it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.97it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.96it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.93it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.96it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.95it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.90it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.94it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  5.18it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  5.02it/s]
+Evaluated 400 examples
+[2026-07-21 07:51:05,505][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:07,  1.56it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:01<00:06,  1.63it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:01<00:05,  1.73it/s]
Calculating text similarity:  31%|███       | 4/13 [00:02<00:04,  1.85it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:02<00:04,  1.87it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:03<00:04,  1.70it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:04<00:03,  1.62it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:04<00:03,  1.63it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:05<00:02,  1.65it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:05<00:01,  1.67it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:06<00:01,  1.58it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:07<00:00,  1.56it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:07<00:00,  1.70it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:07<00:00,  1.67it/s]
+Evaluated 400 examples
+[2026-07-21 07:51:13,675][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.90it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.92it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.89it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.94it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.97it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.90it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.87it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.82it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.82it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.83it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.81it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.83it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  5.09it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.92it/s]
+Evaluated 400 examples
+[2026-07-21 07:51:16,596][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:11,  1.02it/s]
Calculating loss:  15%|█▌        | 2/13 [00:01<00:10,  1.02it/s]
Calculating loss:  23%|██▎       | 3/13 [00:02<00:09,  1.01it/s]
Calculating loss:  31%|███       | 4/13 [00:03<00:08,  1.02it/s]
Calculating loss:  38%|███▊      | 5/13 [00:04<00:07,  1.02it/s]
Calculating loss:  46%|████▌     | 6/13 [00:05<00:06,  1.01it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:06<00:05,  1.00it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:07<00:05,  1.00s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:08<00:04,  1.01s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:09<00:03,  1.00s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:10<00:02,  1.01s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:11<00:01,  1.01s/it]
Calculating loss: 100%|██████████| 13/13 [00:12<00:00,  1.04it/s]
Calculating loss: 100%|██████████| 13/13 [00:12<00:00,  1.01it/s]
+Evaluated 400 examples
+[2026-07-21 07:51:29,441][metrics][INFO] - Evaluating retain_Truth_Ratio
+[2026-07-21 07:51:30,018][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.40it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.83it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  7.97it/s]
+Evaluated 100 examples
+[2026-07-21 07:51:30,782][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.32it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.62it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.22it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.64it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.67it/s]
+Evaluated 100 examples
+[2026-07-21 07:51:32,283][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+[2026-07-21 07:51:32,543][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:01,  2.62it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  2.93it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:01<00:00,  3.02it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  3.49it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  3.25it/s]
+Evaluated 100 examples
+[2026-07-21 07:51:33,776][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-21 07:51:33,776][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-21 07:51:33,776][metrics][INFO] - Evaluating ra_Truth_Ratio
+[2026-07-21 07:51:34,039][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  7.24it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.40it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  8.00it/s]
+Evaluated 117 examples
+[2026-07-21 07:51:34,800][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.47it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.61it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.21it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.70it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.71it/s]
+Evaluated 117 examples
+[2026-07-21 07:51:36,277][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+[2026-07-21 07:51:36,535][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:01,  2.29it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  2.38it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:01<00:00,  2.28it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  2.32it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  2.32it/s]
+Evaluated 117 examples
+[2026-07-21 07:51:38,263][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-21 07:51:38,263][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-21 07:51:38,263][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-21 07:51:38,263][metrics][INFO] - Evaluating model_utility
+[2026-07-21 07:51:38,264][evaluator][INFO] - Result for metric model_utility:	0.5621805155060373
+[2026-07-21 07:51:38,799][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-21 07:51:38,814][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.07it/s]
 31%|███       | 4/13 [00:00<00:00, 11.25it/s]
 46%|████▌     | 6/13 [00:00<00:00, 11.55it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 11.73it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 11.97it/s]
 92%|█████████▏| 12/13 [00:01<00:00, 11.71it/s]
100%|██████████| 13/13 [00:01<00:00, 12.24it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 11.49it/s]
 31%|███       | 4/13 [00:00<00:00, 11.33it/s]
 46%|████▌     | 6/13 [00:00<00:00, 11.62it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 11.69it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 11.71it/s]
 92%|█████████▏| 12/13 [00:01<00:00, 11.84it/s]
100%|██████████| 13/13 [00:01<00:00, 12.12it/s]
+[2026-07-21 07:51:40,954][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-21 07:51:40,968][metrics][INFO] - Evaluating privleak
+[2026-07-21 07:51:40,969][evaluator][INFO] - Result for metric privleak:	-98.67081045556141
+[2026-07-21 07:51:41,263][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:01, 10.80it/s]
Calculating ES:  31%|███       | 4/13 [00:00<00:00,  9.63it/s]
Calculating ES:  46%|████▌     | 6/13 [00:00<00:00, 10.20it/s]
Calculating ES:  62%|██████▏   | 8/13 [00:00<00:00, 10.46it/s]
Calculating ES:  77%|███████▋  | 10/13 [00:00<00:00, 10.41it/s]
Calculating ES:  92%|█████████▏| 12/13 [00:01<00:00, 10.24it/s]
Calculating ES: 100%|██████████| 13/13 [00:01<00:00, 10.66it/s]
+Evaluated 400 examples
+[2026-07-21 07:51:42,482][evaluator][INFO] - Result for metric extraction_strength:	0.3781519412748065
diff --git a/logs/eval_IdkNLL_lr4e-05_alpha10_epoch5.log b/logs/eval_IdkNLL_lr4e-05_alpha10_epoch5.log
new file mode 100644
index 0000000000000000000000000000000000000000..b9d684093b7a0321f6451a1180b3a2c10490f1dc
--- /dev/null
+++ b/logs/eval_IdkNLL_lr4e-05_alpha10_epoch5.log
@@ -0,0 +1,78 @@
+/opt/conda/envs/copygen/lib/python3.12/site-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.6.3) or chardet (6.0.0.post1)/charset_normalizer (3.4.4) doesn't match a supported version!
+  warnings.warn(
+[2026-07-22 17:18:02,596][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/IdkNLL_lr4e-05_alpha10_epoch5
+[2026-07-22 17:18:02,596][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-22 17:18:02,596][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/IdkNLL_lr4e-05_alpha10_epoch5/TOFU_EVAL.json
+[2026-07-22 17:18:02,597][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/IdkNLL_lr4e-05_alpha10_epoch5/TOFU_SUMMARY.json
+[2026-07-22 17:18:04,362][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:05,  2.31it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:03,  3.21it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  3.59it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.84it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.04it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.14it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.23it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:02<00:01,  4.35it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.39it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.43it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.41it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.34it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.52it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.14it/s]
+Evaluated 400 examples
+[2026-07-22 17:18:08,168][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:13,  1.11s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:12,  1.11s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:11,  1.14s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:10,  1.14s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:09,  1.13s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.13s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.13s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.11s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:10<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:11<00:03,  1.10s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:12<00:02,  1.11s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.12s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.06s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.10s/it]
+Evaluated 400 examples
+[2026-07-22 17:18:22,511][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-22 17:18:22,512][evaluator][INFO] - Result for metric forget_truth_ratio:	0.5229820334714074
+[2026-07-22 17:18:22,516][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-22 17:18:22,516][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:18:22,524][metrics][INFO] - Evaluating forget_quality
+[2026-07-22 17:18:22,525][evaluator][INFO] - Result for metric forget_quality:	7.975373070421785e-17
+[2026-07-22 17:18:22,882][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.39it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.44it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.25it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:02,  4.26it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.29it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.35it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.37it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.49it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.55it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.55it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.53it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.51it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.76it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.51it/s]
+Evaluated 400 examples
+[2026-07-22 17:18:25,768][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.6660107421875
+[2026-07-22 17:18:26,079][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:04,  2.90it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:03,  3.50it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:02,  3.50it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:02,  3.76it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:02,  3.82it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:01<00:01,  4.10it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:01<00:01,  4.17it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  4.43it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:02<00:00,  4.84it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:02<00:00,  4.37it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:02<00:00,  3.96it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:03<00:00,  3.85it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  4.20it/s]
+Evaluated 400 examples
+[2026-07-22 17:18:29,177][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.08163743550245471
+[2026-07-22 17:18:29,433][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.62it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.68it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.68it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.71it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.66it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.61it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.61it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.55it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.59it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.59it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.51it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.53it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.74it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.64it/s]
+Evaluated 400 examples
+[2026-07-22 17:18:32,521][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:03,  3.09it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:03,  3.46it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:02,  3.61it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:02,  3.80it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:02,  3.71it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:01<00:01,  3.51it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:01<00:01,  3.43it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  3.40it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:02<00:01,  3.48it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:02<00:00,  3.27it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:03<00:00,  3.19it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:03<00:00,  3.14it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.42it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.41it/s]
+Evaluated 400 examples
+[2026-07-22 17:18:36,610][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.59it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.60it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.56it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.62it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.65it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.58it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.54it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.49it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.49it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.50it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.48it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.50it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.74it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.59it/s]
+Evaluated 400 examples
+[2026-07-22 17:18:39,732][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:13,  1.10s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:11,  1.08s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:10,  1.08s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:09,  1.07s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:08,  1.06s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.08s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.09s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.10s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:09<00:04,  1.11s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:10<00:03,  1.10s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:12<00:02,  1.11s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.10s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.05s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.08s/it]
+Evaluated 400 examples
+[2026-07-22 17:18:53,769][metrics][INFO] - Evaluating retain_Truth_Ratio
+[2026-07-22 17:18:54,165][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.17it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.58it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  7.49it/s]
+Evaluated 100 examples
+[2026-07-22 17:18:55,096][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.39it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.49it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.09it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.45it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.50it/s]
+Evaluated 100 examples
+[2026-07-22 17:18:56,697][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+[2026-07-22 17:18:57,023][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  3.49it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  4.72it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  4.80it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  5.69it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  5.15it/s]
+Evaluated 100 examples
+[2026-07-22 17:18:57,800][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-22 17:18:57,800][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:18:57,800][metrics][INFO] - Evaluating ra_Truth_Ratio
+[2026-07-22 17:18:58,036][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.12it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.62it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  8.40it/s]
+Evaluated 117 examples
+[2026-07-22 17:18:58,830][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.83it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.52it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.11it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.57it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.59it/s]
+Evaluated 117 examples
+[2026-07-22 17:19:00,373][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+[2026-07-22 17:19:00,638][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  4.06it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  4.43it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  4.01it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  3.97it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  4.03it/s]
+Evaluated 117 examples
+[2026-07-22 17:19:01,631][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-22 17:19:01,631][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:19:01,631][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-22 17:19:01,631][metrics][INFO] - Evaluating model_utility
+[2026-07-22 17:19:01,632][evaluator][INFO] - Result for metric model_utility:	0.5490092971263351
+[2026-07-22 17:19:02,234][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:19:02,242][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.67it/s]
 31%|███       | 4/13 [00:00<00:00, 12.23it/s]
 46%|████▌     | 6/13 [00:00<00:00, 12.77it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.11it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.44it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.14it/s]
100%|██████████| 13/13 [00:00<00:00, 13.62it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 12.65it/s]
 31%|███       | 4/13 [00:00<00:00, 12.60it/s]
 46%|████▌     | 6/13 [00:00<00:00, 13.05it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.08it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.17it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.11it/s]
100%|██████████| 13/13 [00:00<00:00, 13.47it/s]
+[2026-07-22 17:19:04,164][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:19:04,171][metrics][INFO] - Evaluating privleak
+[2026-07-22 17:19:04,171][evaluator][INFO] - Result for metric privleak:	-98.57868840904965
+[2026-07-22 17:19:04,556][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:01, 10.93it/s]
Calculating ES:  31%|███       | 4/13 [00:00<00:00, 10.18it/s]
Calculating ES:  46%|████▌     | 6/13 [00:00<00:00, 10.67it/s]
Calculating ES:  62%|██████▏   | 8/13 [00:00<00:00, 11.12it/s]
Calculating ES:  77%|███████▋  | 10/13 [00:00<00:00, 11.27it/s]
Calculating ES:  92%|█████████▏| 12/13 [00:01<00:00, 11.12it/s]
Calculating ES: 100%|██████████| 13/13 [00:01<00:00, 11.47it/s]
+Evaluated 400 examples
+[2026-07-22 17:19:05,690][evaluator][INFO] - Result for metric extraction_strength:	0.3448394188954561
diff --git a/logs/eval_IdkNLL_lr4e-05_alpha1_epoch10.log b/logs/eval_IdkNLL_lr4e-05_alpha1_epoch10.log
new file mode 100644
index 0000000000000000000000000000000000000000..5bb3bd10213405ac5a7faa8f6ca06228d6616dec
--- /dev/null
+++ b/logs/eval_IdkNLL_lr4e-05_alpha1_epoch10.log
@@ -0,0 +1,78 @@
+/opt/conda/envs/copygen/lib/python3.12/site-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.6.3) or chardet (6.0.0.post1)/charset_normalizer (3.4.4) doesn't match a supported version!
+  warnings.warn(
+[2026-07-22 17:11:18,673][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/IdkNLL_lr4e-05_alpha1_epoch10
+[2026-07-22 17:11:18,674][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-22 17:11:18,674][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/IdkNLL_lr4e-05_alpha1_epoch10/TOFU_EVAL.json
+[2026-07-22 17:11:18,674][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/IdkNLL_lr4e-05_alpha1_epoch10/TOFU_SUMMARY.json
+[2026-07-22 17:11:19,386][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:05,  2.17it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:03,  3.08it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  3.47it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.70it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:02,  3.90it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.00it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.11it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:02<00:01,  4.23it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.30it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.36it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.35it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:03<00:00,  4.35it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.56it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.06it/s]
+Evaluated 400 examples
+[2026-07-22 17:11:22,928][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:13,  1.12s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:12,  1.12s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:11,  1.15s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:10,  1.15s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:09,  1.14s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.14s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.13s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:09<00:05,  1.11s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:10<00:04,  1.11s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:11<00:03,  1.11s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:12<00:02,  1.12s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.12s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.07s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.11s/it]
+Evaluated 400 examples
+[2026-07-22 17:11:37,368][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-22 17:11:37,368][evaluator][INFO] - Result for metric forget_truth_ratio:	0.5215733989666866
+[2026-07-22 17:11:37,373][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-22 17:11:37,374][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:11:37,381][metrics][INFO] - Evaluating forget_quality
+[2026-07-22 17:11:37,382][evaluator][INFO] - Result for metric forget_quality:	7.975373070421785e-17
+[2026-07-22 17:11:37,739][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.36it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.36it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.26it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:02,  4.32it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.37it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.43it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.43it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.53it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.58it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.57it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.53it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.51it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.75it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.52it/s]
+Evaluated 400 examples
+[2026-07-22 17:11:40,613][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.587607421875
+[2026-07-22 17:11:40,923][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:01,  9.34it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:00, 10.50it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:00<00:00, 11.09it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:00<00:00, 11.61it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:00<00:00, 11.83it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:00<00:00, 11.80it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:01<00:00, 10.18it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:01<00:00, 10.78it/s]
+Evaluated 400 examples
+[2026-07-22 17:11:42,129][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.0071564874892455456
+[2026-07-22 17:11:42,380][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.62it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.68it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.69it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.71it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.71it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.64it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.63it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.60it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.63it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.62it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.57it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.59it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.82it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.68it/s]
+Evaluated 400 examples
+[2026-07-22 17:11:45,392][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:03,  3.57it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:03,  3.03it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:01<00:03,  2.87it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:02,  3.08it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:02,  3.28it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:01<00:02,  3.46it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:02<00:01,  3.43it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  3.28it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:02<00:01,  3.23it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:03<00:00,  3.36it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:03<00:00,  3.18it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:03<00:00,  3.30it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.58it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.33it/s]
+Evaluated 400 examples
+[2026-07-22 17:11:49,551][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.63it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.61it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.57it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.57it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.61it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.54it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.51it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.46it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.47it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.49it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.46it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.49it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.73it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.57it/s]
+Evaluated 400 examples
+[2026-07-22 17:11:52,631][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:12,  1.07s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:11,  1.08s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:10,  1.08s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:09,  1.07s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:08,  1.06s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.08s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.09s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.09s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:09<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:10<00:03,  1.10s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:11<00:02,  1.11s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.10s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.05s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.08s/it]
+Evaluated 400 examples
+[2026-07-22 17:12:06,651][metrics][INFO] - Evaluating retain_Truth_Ratio
+[2026-07-22 17:12:07,128][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.14it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.53it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  7.56it/s]
+Evaluated 100 examples
+[2026-07-22 17:12:07,914][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.33it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.48it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.07it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.45it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.49it/s]
+Evaluated 100 examples
+[2026-07-22 17:12:09,519][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+[2026-07-22 17:12:09,747][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  4.78it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  5.50it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  5.24it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  6.27it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  5.84it/s]
+Evaluated 100 examples
+[2026-07-22 17:12:10,432][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-22 17:12:10,433][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:12:10,433][metrics][INFO] - Evaluating ra_Truth_Ratio
+[2026-07-22 17:12:10,706][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  7.93it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.50it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  8.25it/s]
+Evaluated 117 examples
+[2026-07-22 17:12:11,472][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.81it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.51it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.10it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.57it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.59it/s]
+Evaluated 117 examples
+[2026-07-22 17:12:13,017][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+[2026-07-22 17:12:13,262][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  4.89it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  5.33it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  4.39it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  4.53it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  4.62it/s]
+Evaluated 117 examples
+[2026-07-22 17:12:14,129][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-22 17:12:14,129][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:12:14,129][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-22 17:12:14,129][metrics][INFO] - Evaluating model_utility
+[2026-07-22 17:12:14,129][evaluator][INFO] - Result for metric model_utility:	0.508652343560777
+[2026-07-22 17:12:14,623][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:12:14,630][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.67it/s]
 31%|███       | 4/13 [00:00<00:00, 12.16it/s]
 46%|████▌     | 6/13 [00:00<00:00, 12.85it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.23it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.51it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.30it/s]
100%|██████████| 13/13 [00:00<00:00, 13.75it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.12it/s]
 31%|███       | 4/13 [00:00<00:00, 12.92it/s]
 46%|████▌     | 6/13 [00:00<00:00, 13.33it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.33it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.35it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.11it/s]
100%|██████████| 13/13 [00:00<00:00, 13.56it/s]
+[2026-07-22 17:12:16,537][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:12:16,544][metrics][INFO] - Evaluating privleak
+[2026-07-22 17:12:16,544][evaluator][INFO] - Result for metric privleak:	-97.12599459867167
+[2026-07-22 17:12:16,797][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:00, 11.23it/s]
Calculating ES:  31%|███       | 4/13 [00:00<00:00, 10.20it/s]
Calculating ES:  46%|████▌     | 6/13 [00:00<00:00, 10.14it/s]
Calculating ES:  62%|██████▏   | 8/13 [00:00<00:00,  9.97it/s]
Calculating ES:  77%|███████▋  | 10/13 [00:00<00:00, 10.20it/s]
Calculating ES:  92%|█████████▏| 12/13 [00:01<00:00, 10.14it/s]
Calculating ES: 100%|██████████| 13/13 [00:01<00:00, 10.63it/s]
+Evaluated 400 examples
+[2026-07-22 17:12:18,021][evaluator][INFO] - Result for metric extraction_strength:	0.29804959133370734
diff --git a/logs/eval_IdkNLL_lr4e-05_alpha1_epoch5.log b/logs/eval_IdkNLL_lr4e-05_alpha1_epoch5.log
new file mode 100644
index 0000000000000000000000000000000000000000..d2b9b48db9d9ad72fc38bbc8f5bc8f7184bbdedf
--- /dev/null
+++ b/logs/eval_IdkNLL_lr4e-05_alpha1_epoch5.log
@@ -0,0 +1,78 @@
+/opt/conda/envs/copygen/lib/python3.12/site-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.6.3) or chardet (6.0.0.post1)/charset_normalizer (3.4.4) doesn't match a supported version!
+  warnings.warn(
+[2026-07-22 17:19:36,919][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/IdkNLL_lr4e-05_alpha1_epoch5
+[2026-07-22 17:19:36,920][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-22 17:19:36,920][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/IdkNLL_lr4e-05_alpha1_epoch5/TOFU_EVAL.json
+[2026-07-22 17:19:36,920][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/IdkNLL_lr4e-05_alpha1_epoch5/TOFU_SUMMARY.json
+[2026-07-22 17:19:37,585][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:05,  2.27it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:03,  3.17it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  3.57it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.79it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.01it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.10it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.17it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:02<00:01,  4.28it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.32it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.38it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.38it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.37it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.59it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.12it/s]
+Evaluated 400 examples
+[2026-07-22 17:19:41,065][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:13,  1.11s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:12,  1.11s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:11,  1.14s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:10,  1.15s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:09,  1.13s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.13s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.13s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.11s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:10<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:11<00:03,  1.10s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:12<00:02,  1.11s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.11s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.06s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.10s/it]
+Evaluated 400 examples
+[2026-07-22 17:19:55,423][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-22 17:19:55,424][evaluator][INFO] - Result for metric forget_truth_ratio:	0.5261941641418377
+[2026-07-22 17:19:55,428][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-22 17:19:55,428][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:19:55,436][metrics][INFO] - Evaluating forget_quality
+[2026-07-22 17:19:55,436][evaluator][INFO] - Result for metric forget_quality:	5.137999089404319e-16
+[2026-07-22 17:19:55,792][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.26it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.31it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.20it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:02,  4.27it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.34it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.41it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.42it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.52it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.57it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.57it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.54it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.52it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.76it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.51it/s]
+Evaluated 400 examples
+[2026-07-22 17:19:58,673][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.59710693359375
+[2026-07-22 17:19:58,927][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:01,  9.22it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:00, 10.52it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:00<00:00, 11.10it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:00<00:00, 11.40it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:00<00:00, 11.37it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:00<00:00, 11.24it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:01<00:00, 11.01it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:01<00:00, 11.05it/s]
+Evaluated 400 examples
+[2026-07-22 17:20:00,104][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.004809031889425527
+[2026-07-22 17:20:00,391][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.51it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.64it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.68it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.71it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.71it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.65it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.64it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.60it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.63it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.63it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.57it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.61it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.83it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.68it/s]
+Evaluated 400 examples
+[2026-07-22 17:20:03,434][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:03,  3.52it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:03,  3.44it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:02,  3.41it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:02,  3.35it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:02,  3.41it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:01<00:01,  3.51it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:02<00:01,  3.41it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  3.45it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:02<00:01,  3.28it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:02<00:00,  3.40it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:03<00:00,  3.33it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:03<00:00,  3.43it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.32it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.38it/s]
+Evaluated 400 examples
+[2026-07-22 17:20:07,516][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.61it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.62it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.60it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.66it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.69it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.57it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.54it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.49it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.49it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.51it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.48it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.51it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.75it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.60it/s]
+Evaluated 400 examples
+[2026-07-22 17:20:10,603][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:12,  1.08s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:11,  1.08s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:10,  1.09s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:09,  1.07s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:08,  1.06s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.07s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.09s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.10s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:09<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:10<00:03,  1.10s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:12<00:02,  1.11s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.10s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.05s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.08s/it]
+Evaluated 400 examples
+[2026-07-22 17:20:24,644][metrics][INFO] - Evaluating retain_Truth_Ratio
+[2026-07-22 17:20:25,166][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.01it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.49it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  7.42it/s]
+Evaluated 100 examples
+[2026-07-22 17:20:25,953][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.50it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.50it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.09it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.43it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.49it/s]
+Evaluated 100 examples
+[2026-07-22 17:20:27,562][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+[2026-07-22 17:20:27,808][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  5.69it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  5.85it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  6.07it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  6.81it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  6.45it/s]
+Evaluated 100 examples
+[2026-07-22 17:20:28,430][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-22 17:20:28,430][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:20:28,430][metrics][INFO] - Evaluating ra_Truth_Ratio
+[2026-07-22 17:20:28,699][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  7.87it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.39it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  8.09it/s]
+Evaluated 117 examples
+[2026-07-22 17:20:29,429][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.59it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.47it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.09it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.57it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.58it/s]
+Evaluated 117 examples
+[2026-07-22 17:20:30,982][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+[2026-07-22 17:20:31,211][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  4.70it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  4.77it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  4.01it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  4.27it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  4.31it/s]
+Evaluated 117 examples
+[2026-07-22 17:20:32,139][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-22 17:20:32,139][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:20:32,139][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-22 17:20:32,139][metrics][INFO] - Evaluating model_utility
+[2026-07-22 17:20:32,140][evaluator][INFO] - Result for metric model_utility:	0.4534620375113887
+[2026-07-22 17:20:32,692][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:20:32,700][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 14.22it/s]
 31%|███       | 4/13 [00:00<00:00, 12.93it/s]
 46%|████▌     | 6/13 [00:00<00:00, 13.59it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.93it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 14.27it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.98it/s]
100%|██████████| 13/13 [00:00<00:00, 14.49it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.20it/s]
 31%|███       | 4/13 [00:00<00:00, 13.26it/s]
 46%|████▌     | 6/13 [00:00<00:00, 13.72it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.72it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.81it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.91it/s]
100%|██████████| 13/13 [00:00<00:00, 14.20it/s]
+[2026-07-22 17:20:34,514][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:20:34,521][metrics][INFO] - Evaluating privleak
+[2026-07-22 17:20:34,522][evaluator][INFO] - Result for metric privleak:	-97.43273064365043
+[2026-07-22 17:20:34,869][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:00, 11.56it/s]
Calculating ES:  31%|███       | 4/13 [00:00<00:00, 10.75it/s]
Calculating ES:  46%|████▌     | 6/13 [00:00<00:00, 11.48it/s]
Calculating ES:  62%|██████▏   | 8/13 [00:00<00:00, 11.75it/s]
Calculating ES:  77%|███████▋  | 10/13 [00:00<00:00, 11.72it/s]
Calculating ES:  92%|█████████▏| 12/13 [00:01<00:00, 11.37it/s]
Calculating ES: 100%|██████████| 13/13 [00:01<00:00, 11.92it/s]
+Evaluated 400 examples
+[2026-07-22 17:20:35,961][evaluator][INFO] - Result for metric extraction_strength:	0.29292188014521775
diff --git a/logs/eval_IdkNLL_lr4e-05_alpha2_epoch10.log b/logs/eval_IdkNLL_lr4e-05_alpha2_epoch10.log
new file mode 100644
index 0000000000000000000000000000000000000000..2ec95242669353bf721f1cead11b79d8b0ae7506
--- /dev/null
+++ b/logs/eval_IdkNLL_lr4e-05_alpha2_epoch10.log
@@ -0,0 +1,76 @@
+[2026-07-21 07:50:29,562][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/IdkNLL_lr4e-05_alpha2_epoch10
+[2026-07-21 07:50:29,563][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-21 07:50:29,564][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/IdkNLL_lr4e-05_alpha2_epoch10/TOFU_EVAL.json
+[2026-07-21 07:50:29,564][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/IdkNLL_lr4e-05_alpha2_epoch10/TOFU_SUMMARY.json
+[2026-07-21 07:50:30,301][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:08,  1.49it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:04,  2.50it/s]
Calculating loss:  23%|██▎       | 3/13 [00:01<00:03,  3.15it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.61it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:02,  3.96it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.19it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.37it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:02<00:01,  4.53it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.61it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.66it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.67it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.67it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.82it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.08it/s]
+Evaluated 400 examples
+[2026-07-21 07:50:33,815][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:12,  1.03s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:11,  1.02s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:10,  1.04s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:09,  1.04s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:08,  1.03s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.03s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.02s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.01s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:09<00:04,  1.01s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:10<00:03,  1.01s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:11<00:02,  1.02s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:12<00:01,  1.02s/it]
Calculating loss: 100%|██████████| 13/13 [00:13<00:00,  1.03it/s]
Calculating loss: 100%|██████████| 13/13 [00:13<00:00,  1.01s/it]
+Evaluated 400 examples
+[2026-07-21 07:50:46,933][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-21 07:50:46,933][evaluator][INFO] - Result for metric forget_truth_ratio:	0.5161614943622759
+[2026-07-21 07:50:46,941][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-21 07:50:46,942][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-21 07:50:46,958][metrics][INFO] - Evaluating forget_quality
+[2026-07-21 07:50:46,960][evaluator][INFO] - Result for metric forget_quality:	7.975373070421785e-17
+[2026-07-21 07:50:47,384][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.81it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.78it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.67it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.71it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.73it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.78it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.78it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.86it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.89it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.87it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.85it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.84it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  5.10it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.87it/s]
+Evaluated 400 examples
+[2026-07-21 07:50:50,051][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.60478271484375
+[2026-07-21 07:50:50,320][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:02,  5.05it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:01,  6.00it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:01,  6.29it/s]
Calculating text similarity:  31%|███       | 4/13 [00:00<00:01,  6.51it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:00<00:01,  6.56it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:00<00:01,  6.74it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:01<00:00,  6.87it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:01<00:00,  6.97it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:01<00:00,  4.86it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:01<00:00,  5.35it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:02<00:00,  4.15it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:02<00:00,  4.73it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:02<00:00,  5.43it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:02<00:00,  5.59it/s]
+Evaluated 400 examples
+[2026-07-21 07:50:52,646][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.011540083002540306
+[2026-07-21 07:50:52,921][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.97it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  5.03it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:01,  5.04it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  5.06it/s]
Calculating loss:  38%|███▊      | 5/13 [00:00<00:01,  5.06it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.99it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.98it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.95it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.98it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.97it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.92it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.96it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  5.20it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  5.04it/s]
+Evaluated 400 examples
+[2026-07-21 07:50:55,768][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:06,  1.99it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:01<00:06,  1.77it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:01<00:05,  1.83it/s]
Calculating text similarity:  31%|███       | 4/13 [00:02<00:04,  1.93it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:02<00:03,  2.03it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:03<00:03,  1.95it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:03<00:03,  1.81it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:04<00:02,  1.78it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:04<00:02,  1.81it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:05<00:01,  1.84it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:06<00:01,  1.69it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:06<00:00,  1.61it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:07<00:00,  1.75it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:07<00:00,  1.80it/s]
+Evaluated 400 examples
+[2026-07-21 07:51:03,352][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.92it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.94it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.91it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.96it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.99it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.91it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.89it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.84it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.83it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.84it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.82it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.84it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  5.10it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.94it/s]
+Evaluated 400 examples
+[2026-07-21 07:51:06,243][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:11,  1.02it/s]
Calculating loss:  15%|█▌        | 2/13 [00:01<00:10,  1.02it/s]
Calculating loss:  23%|██▎       | 3/13 [00:02<00:09,  1.01it/s]
Calculating loss:  31%|███       | 4/13 [00:03<00:08,  1.02it/s]
Calculating loss:  38%|███▊      | 5/13 [00:04<00:07,  1.03it/s]
Calculating loss:  46%|████▌     | 6/13 [00:05<00:06,  1.01it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:06<00:05,  1.00it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:07<00:05,  1.00s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:08<00:04,  1.00s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:09<00:03,  1.00s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:10<00:02,  1.01s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:11<00:01,  1.00s/it]
Calculating loss: 100%|██████████| 13/13 [00:12<00:00,  1.04it/s]
Calculating loss: 100%|██████████| 13/13 [00:12<00:00,  1.02it/s]
+Evaluated 400 examples
+[2026-07-21 07:51:19,046][metrics][INFO] - Evaluating retain_Truth_Ratio
+[2026-07-21 07:51:19,411][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.42it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.85it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  7.99it/s]
+Evaluated 100 examples
+[2026-07-21 07:51:20,232][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.38it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.63it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.22it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.64it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.67it/s]
+Evaluated 100 examples
+[2026-07-21 07:51:21,730][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+[2026-07-21 07:51:22,002][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:01,  1.81it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:01<00:00,  2.04it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:01<00:00,  2.44it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  2.98it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  2.61it/s]
+Evaluated 100 examples
+[2026-07-21 07:51:23,537][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-21 07:51:23,537][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-21 07:51:23,537][metrics][INFO] - Evaluating ra_Truth_Ratio
+[2026-07-21 07:51:23,816][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.37it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.89it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  8.66it/s]
+Evaluated 117 examples
+[2026-07-21 07:51:24,530][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.58it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.62it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.22it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.72it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.72it/s]
+Evaluated 117 examples
+[2026-07-21 07:51:25,999][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+[2026-07-21 07:51:26,277][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:01,  2.34it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  2.39it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:01<00:00,  1.94it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  2.11it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  2.13it/s]
+Evaluated 117 examples
+[2026-07-21 07:51:28,159][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-21 07:51:28,159][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-21 07:51:28,159][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-21 07:51:28,160][metrics][INFO] - Evaluating model_utility
+[2026-07-21 07:51:28,161][evaluator][INFO] - Result for metric model_utility:	0.5427105442935864
+[2026-07-21 07:51:28,731][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-21 07:51:28,746][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.46it/s]
 31%|███       | 4/13 [00:00<00:00, 11.58it/s]
 46%|████▌     | 6/13 [00:00<00:00, 11.87it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 12.01it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 12.23it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 11.96it/s]
100%|██████████| 13/13 [00:01<00:00, 12.52it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 11.72it/s]
 31%|███       | 4/13 [00:00<00:00, 11.59it/s]
 46%|████▌     | 6/13 [00:00<00:00, 11.88it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 11.92it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 11.96it/s]
 92%|█████████▏| 12/13 [00:01<00:00, 12.07it/s]
100%|██████████| 13/13 [00:01<00:00, 12.37it/s]
+[2026-07-21 07:51:30,839][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-21 07:51:30,853][metrics][INFO] - Evaluating privleak
+[2026-07-21 07:51:30,853][evaluator][INFO] - Result for metric privleak:	-97.62304871776266
+[2026-07-21 07:51:31,136][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:01, 10.93it/s]
Calculating ES:  31%|███       | 4/13 [00:00<00:00,  9.71it/s]
Calculating ES:  46%|████▌     | 6/13 [00:00<00:00, 10.23it/s]
Calculating ES:  62%|██████▏   | 8/13 [00:00<00:00, 10.48it/s]
Calculating ES:  77%|███████▋  | 10/13 [00:00<00:00, 10.55it/s]
Calculating ES:  92%|█████████▏| 12/13 [00:01<00:00, 10.35it/s]
Calculating ES: 100%|██████████| 13/13 [00:01<00:00, 10.76it/s]
+Evaluated 400 examples
+[2026-07-21 07:51:32,345][evaluator][INFO] - Result for metric extraction_strength:	0.31056017504923966
diff --git a/logs/eval_IdkNLL_lr4e-05_alpha2_epoch5.log b/logs/eval_IdkNLL_lr4e-05_alpha2_epoch5.log
new file mode 100644
index 0000000000000000000000000000000000000000..e1e6e4e730c3f0c21037b68df9d16eab41222196
--- /dev/null
+++ b/logs/eval_IdkNLL_lr4e-05_alpha2_epoch5.log
@@ -0,0 +1,78 @@
+/opt/conda/envs/copygen/lib/python3.12/site-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.6.3) or chardet (6.0.0.post1)/charset_normalizer (3.4.4) doesn't match a supported version!
+  warnings.warn(
+[2026-07-22 17:12:44,681][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/IdkNLL_lr4e-05_alpha2_epoch5
+[2026-07-22 17:12:44,682][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-22 17:12:44,682][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/IdkNLL_lr4e-05_alpha2_epoch5/TOFU_EVAL.json
+[2026-07-22 17:12:44,682][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/IdkNLL_lr4e-05_alpha2_epoch5/TOFU_SUMMARY.json
+[2026-07-22 17:12:45,331][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:04,  2.42it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:03,  3.30it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  3.66it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.89it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.08it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.18it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.27it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:02<00:01,  4.38it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.42it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.44it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.41it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.39it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.55it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.18it/s]
+Evaluated 400 examples
+[2026-07-22 17:12:48,779][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:13,  1.13s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:12,  1.13s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:11,  1.16s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:10,  1.16s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:09,  1.14s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.13s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.13s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:09<00:05,  1.11s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:10<00:04,  1.11s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:11<00:03,  1.11s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:12<00:02,  1.12s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.13s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.07s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.11s/it]
+Evaluated 400 examples
+[2026-07-22 17:13:03,234][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-22 17:13:03,235][evaluator][INFO] - Result for metric forget_truth_ratio:	0.5212455071811525
+[2026-07-22 17:13:03,239][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-22 17:13:03,239][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:13:03,246][metrics][INFO] - Evaluating forget_quality
+[2026-07-22 17:13:03,247][evaluator][INFO] - Result for metric forget_quality:	2.776495661774257e-16
+[2026-07-22 17:13:03,668][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.46it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.47it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.32it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:02,  4.34it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.38it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.43it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.42it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.50it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.50it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.50it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.47it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.46it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.69it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.50it/s]
+Evaluated 400 examples
+[2026-07-22 17:13:06,556][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.6036669921875
+[2026-07-22 17:13:06,861][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:02,  4.68it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:02,  5.18it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:01,  6.46it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:00<00:01,  7.98it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:00<00:00,  8.47it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:00<00:00,  8.82it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:01<00:00,  8.87it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:01<00:00,  7.24it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:01<00:00,  5.74it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:01<00:00,  6.44it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:01<00:00,  8.07it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:01<00:00,  7.37it/s]
+Evaluated 400 examples
+[2026-07-22 17:13:08,626][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.013236566883255176
+[2026-07-22 17:13:08,879][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.64it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.58it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.58it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.61it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.61it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.58it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.60it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.58it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.62it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.62it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.58it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.63it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.85it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.66it/s]
+Evaluated 400 examples
+[2026-07-22 17:13:11,954][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:03,  3.49it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:03,  3.58it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:02,  3.66it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:02,  3.53it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:02,  3.13it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:01<00:02,  3.24it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:02<00:01,  3.21it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  3.20it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:02<00:01,  3.11it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:03<00:00,  3.17it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:03<00:00,  3.24it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:03<00:00,  3.46it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.77it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.40it/s]
+Evaluated 400 examples
+[2026-07-22 17:13:16,013][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.59it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.59it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.54it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.61it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.64it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.55it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.51it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.40it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.39it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.40it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.39it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.44it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.70it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.54it/s]
+Evaluated 400 examples
+[2026-07-22 17:13:19,113][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:12,  1.06s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:11,  1.06s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:10,  1.07s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:09,  1.07s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:08,  1.06s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.08s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.10s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.10s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:09<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:10<00:03,  1.09s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:11<00:02,  1.10s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.10s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.05s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.08s/it]
+Evaluated 400 examples
+[2026-07-22 17:13:33,128][metrics][INFO] - Evaluating retain_Truth_Ratio
+[2026-07-22 17:13:33,495][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.13it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.50it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  6.80it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  6.91it/s]
+Evaluated 100 examples
+[2026-07-22 17:13:34,429][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.22it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.49it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.09it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.50it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.53it/s]
+Evaluated 100 examples
+[2026-07-22 17:13:36,009][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+[2026-07-22 17:13:36,383][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  6.57it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  6.34it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  6.28it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  7.12it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  6.82it/s]
+Evaluated 100 examples
+[2026-07-22 17:13:36,970][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-22 17:13:36,970][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:13:36,970][metrics][INFO] - Evaluating ra_Truth_Ratio
+[2026-07-22 17:13:37,209][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.08it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.60it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  8.38it/s]
+Evaluated 117 examples
+[2026-07-22 17:13:37,924][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.85it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.48it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.09it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.56it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.57it/s]
+Evaluated 117 examples
+[2026-07-22 17:13:39,480][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+[2026-07-22 17:13:39,716][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  4.15it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  4.40it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  4.52it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  4.60it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  4.52it/s]
+Evaluated 117 examples
+[2026-07-22 17:13:40,602][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-22 17:13:40,602][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:13:40,602][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-22 17:13:40,602][metrics][INFO] - Evaluating model_utility
+[2026-07-22 17:13:40,603][evaluator][INFO] - Result for metric model_utility:	0.5046823269661426
+[2026-07-22 17:13:41,181][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:13:41,191][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.73it/s]
 31%|███       | 4/13 [00:00<00:00, 12.47it/s]
 46%|████▌     | 6/13 [00:00<00:00, 12.86it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 12.91it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.44it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.21it/s]
100%|██████████| 13/13 [00:00<00:00, 13.65it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 12.65it/s]
 31%|███       | 4/13 [00:00<00:00, 12.58it/s]
 46%|████▌     | 6/13 [00:00<00:00, 12.97it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 12.99it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.15it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.37it/s]
100%|██████████| 13/13 [00:00<00:00, 13.59it/s]
+[2026-07-22 17:13:43,103][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:13:43,109][metrics][INFO] - Evaluating privleak
+[2026-07-22 17:13:43,109][evaluator][INFO] - Result for metric privleak:	-97.46411287927532
+[2026-07-22 17:13:43,360][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:00, 11.82it/s]
Calculating ES:  31%|███       | 4/13 [00:00<00:00, 10.91it/s]
Calculating ES:  46%|████▌     | 6/13 [00:00<00:00, 11.57it/s]
Calculating ES:  62%|██████▏   | 8/13 [00:00<00:00, 11.94it/s]
Calculating ES:  77%|███████▋  | 10/13 [00:00<00:00, 12.16it/s]
Calculating ES:  92%|█████████▏| 12/13 [00:01<00:00, 11.86it/s]
Calculating ES: 100%|██████████| 13/13 [00:01<00:00, 12.28it/s]
+Evaluated 400 examples
+[2026-07-22 17:13:44,419][evaluator][INFO] - Result for metric extraction_strength:	0.3025280251386857
diff --git a/logs/eval_IdkNLL_lr4e-05_alpha5_epoch10.log b/logs/eval_IdkNLL_lr4e-05_alpha5_epoch10.log
new file mode 100644
index 0000000000000000000000000000000000000000..7084783fff40a74bdc63acfa724a8e72b7cf9247
--- /dev/null
+++ b/logs/eval_IdkNLL_lr4e-05_alpha5_epoch10.log
@@ -0,0 +1,78 @@
+/opt/conda/envs/copygen/lib/python3.12/site-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.6.3) or chardet (6.0.0.post1)/charset_normalizer (3.4.4) doesn't match a supported version!
+  warnings.warn(
+[2026-07-22 17:21:08,971][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/IdkNLL_lr4e-05_alpha5_epoch10
+[2026-07-22 17:21:08,972][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-22 17:21:08,972][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/IdkNLL_lr4e-05_alpha5_epoch10/TOFU_EVAL.json
+[2026-07-22 17:21:08,972][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/IdkNLL_lr4e-05_alpha5_epoch10/TOFU_SUMMARY.json
+[2026-07-22 17:21:09,603][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:05,  2.33it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:03,  3.20it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  3.59it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.84it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.02it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.13it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.20it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:02<00:01,  4.29it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.31it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.35it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.35it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.36it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.57it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.13it/s]
+Evaluated 400 examples
+[2026-07-22 17:21:13,107][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:13,  1.11s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:12,  1.11s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:11,  1.14s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:10,  1.14s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:09,  1.13s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.13s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.13s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.11s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:10<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:11<00:03,  1.10s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:12<00:02,  1.11s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.11s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.06s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.10s/it]
+Evaluated 400 examples
+[2026-07-22 17:21:27,448][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-22 17:21:27,448][evaluator][INFO] - Result for metric forget_truth_ratio:	0.5148526803425937
+[2026-07-22 17:21:27,453][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-22 17:21:27,453][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:21:27,461][metrics][INFO] - Evaluating forget_quality
+[2026-07-22 17:21:27,462][evaluator][INFO] - Result for metric forget_quality:	7.975373070421785e-17
+[2026-07-22 17:21:27,830][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.34it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.35it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.24it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:02,  4.31it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.36it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.43it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.43it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.53it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.58it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.57it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.55it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.53it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.77it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.53it/s]
+Evaluated 400 examples
+[2026-07-22 17:21:30,701][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.64310546875
+[2026-07-22 17:21:30,948][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:01,  6.06it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:01,  6.73it/s]
Calculating text similarity:  31%|███       | 4/13 [00:00<00:00,  9.01it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:00<00:01,  7.09it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:00<00:00,  8.75it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:01<00:00,  8.09it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:01<00:00,  7.29it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:01<00:00,  6.55it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:01<00:00,  8.08it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:01<00:00,  7.74it/s]
+Evaluated 400 examples
+[2026-07-22 17:21:32,628][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.021201092885887098
+[2026-07-22 17:21:32,891][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.57it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.67it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.69it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.72it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.72it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.65it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.65it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.60it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.64it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.64it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.58it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.62it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.84it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.69it/s]
+Evaluated 400 examples
+[2026-07-22 17:21:36,057][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:04,  2.89it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:03,  3.30it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:02,  3.47it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:02,  3.60it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:02,  3.64it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:01<00:02,  3.46it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:02<00:01,  3.30it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  3.33it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:02<00:01,  3.39it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:02<00:00,  3.48it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:03<00:00,  3.26it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:03<00:00,  3.21it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.50it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.41it/s]
+Evaluated 400 examples
+[2026-07-22 17:21:40,127][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.56it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.54it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.54it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.61it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.65it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.58it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.54it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.49it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.49it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.51it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.48it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.51it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.74it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.59it/s]
+Evaluated 400 examples
+[2026-07-22 17:21:43,243][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:13,  1.09s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:11,  1.08s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:10,  1.08s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:09,  1.07s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:08,  1.06s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.07s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.09s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.09s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:09<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:10<00:03,  1.10s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:11<00:02,  1.10s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.10s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.05s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.08s/it]
+Evaluated 400 examples
+[2026-07-22 17:21:57,248][metrics][INFO] - Evaluating retain_Truth_Ratio
+[2026-07-22 17:21:57,591][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.14it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.52it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  7.50it/s]
+Evaluated 100 examples
+[2026-07-22 17:21:58,446][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.50it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.49it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.08it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.44it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.49it/s]
+Evaluated 100 examples
+[2026-07-22 17:22:00,051][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+[2026-07-22 17:22:00,291][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  5.29it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  4.33it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  4.54it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  5.45it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  5.10it/s]
+Evaluated 100 examples
+[2026-07-22 17:22:01,076][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-22 17:22:01,076][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:22:01,076][metrics][INFO] - Evaluating ra_Truth_Ratio
+[2026-07-22 17:22:01,317][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.07it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.56it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  8.32it/s]
+Evaluated 117 examples
+[2026-07-22 17:22:02,102][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.73it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.50it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.10it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.57it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.58it/s]
+Evaluated 117 examples
+[2026-07-22 17:22:03,652][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+[2026-07-22 17:22:04,203][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  4.33it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  4.60it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  4.51it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  4.67it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  4.61it/s]
+Evaluated 117 examples
+[2026-07-22 17:22:05,072][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-22 17:22:05,072][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:22:05,072][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-22 17:22:05,072][metrics][INFO] - Evaluating model_utility
+[2026-07-22 17:22:05,072][evaluator][INFO] - Result for metric model_utility:	0.5564634064186886
+[2026-07-22 17:22:05,745][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:22:05,753][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.42it/s]
 31%|███       | 4/13 [00:00<00:00, 12.49it/s]
 46%|████▌     | 6/13 [00:00<00:00, 13.14it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.47it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.84it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.54it/s]
100%|██████████| 13/13 [00:00<00:00, 14.01it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 12.84it/s]
 31%|███       | 4/13 [00:00<00:00, 13.02it/s]
 46%|████▌     | 6/13 [00:00<00:00, 13.50it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.58it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.66it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.60it/s]
100%|██████████| 13/13 [00:00<00:00, 13.97it/s]
+[2026-07-22 17:22:07,615][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:22:07,622][metrics][INFO] - Evaluating privleak
+[2026-07-22 17:22:07,622][evaluator][INFO] - Result for metric privleak:	-98.18894128919214
+[2026-07-22 17:22:07,887][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:00, 11.36it/s]
Calculating ES:  31%|███       | 4/13 [00:00<00:00, 10.05it/s]
Calculating ES:  46%|████▌     | 6/13 [00:00<00:00, 10.66it/s]
Calculating ES:  62%|██████▏   | 8/13 [00:00<00:00, 11.25it/s]
Calculating ES:  77%|███████▋  | 10/13 [00:00<00:00, 11.30it/s]
Calculating ES:  92%|█████████▏| 12/13 [00:01<00:00, 11.24it/s]
Calculating ES: 100%|██████████| 13/13 [00:01<00:00, 11.56it/s]
+Evaluated 400 examples
+[2026-07-22 17:22:09,012][evaluator][INFO] - Result for metric extraction_strength:	0.35020620206680336
diff --git a/logs/eval_IdkNLL_lr4e-05_alpha5_epoch5.log b/logs/eval_IdkNLL_lr4e-05_alpha5_epoch5.log
new file mode 100644
index 0000000000000000000000000000000000000000..5216fd16050ca4f90fc37403e31b597acb41bb19
--- /dev/null
+++ b/logs/eval_IdkNLL_lr4e-05_alpha5_epoch5.log
@@ -0,0 +1,78 @@
+/opt/conda/envs/copygen/lib/python3.12/site-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.6.3) or chardet (6.0.0.post1)/charset_normalizer (3.4.4) doesn't match a supported version!
+  warnings.warn(
+[2026-07-22 17:19:41,479][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/IdkNLL_lr4e-05_alpha5_epoch5
+[2026-07-22 17:19:41,480][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-22 17:19:41,480][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/IdkNLL_lr4e-05_alpha5_epoch5/TOFU_EVAL.json
+[2026-07-22 17:19:41,480][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/IdkNLL_lr4e-05_alpha5_epoch5/TOFU_SUMMARY.json
+[2026-07-22 17:19:42,115][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:05,  2.38it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:03,  3.27it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  3.65it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.88it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.08it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.19it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.25it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:02<00:01,  4.36it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.40it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.44it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.41it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.39it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.60it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.19it/s]
+Evaluated 400 examples
+[2026-07-22 17:19:45,580][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:13,  1.13s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:12,  1.13s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:11,  1.15s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:10,  1.15s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:09,  1.13s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.12s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.12s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.11s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:10<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:11<00:03,  1.11s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:12<00:02,  1.12s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.12s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.06s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.10s/it]
+Evaluated 400 examples
+[2026-07-22 17:19:59,946][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-22 17:19:59,946][evaluator][INFO] - Result for metric forget_truth_ratio:	0.5219291566667374
+[2026-07-22 17:19:59,950][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-22 17:19:59,951][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:19:59,958][metrics][INFO] - Evaluating forget_quality
+[2026-07-22 17:19:59,958][evaluator][INFO] - Result for metric forget_quality:	2.776495661774257e-16
+[2026-07-22 17:20:00,401][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.39it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.45it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.31it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:02,  4.35it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.38it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.43it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.43it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.52it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.56it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.55it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.52it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.50it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.74it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.53it/s]
+Evaluated 400 examples
+[2026-07-22 17:20:03,273][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.6360546875
+[2026-07-22 17:20:03,601][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:02,  4.72it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:02,  5.28it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:02,  3.80it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:01,  4.42it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:01<00:01,  5.22it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:01<00:00,  6.94it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:01<00:00,  7.45it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:01<00:00,  6.52it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:01<00:00,  6.97it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:02<00:00,  6.18it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:02<00:00,  6.15it/s]
+Evaluated 400 examples
+[2026-07-22 17:20:05,715][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.027647814287907826
+[2026-07-22 17:20:06,032][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.68it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.73it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.74it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.76it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.76it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.69it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.68it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.64it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.67it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.66it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.60it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.64it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.86it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.73it/s]
+Evaluated 400 examples
+[2026-07-22 17:20:09,059][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:03,  3.03it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:03,  3.42it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:02,  3.38it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:02,  3.57it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:02,  3.47it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:01<00:02,  3.31it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:02<00:01,  3.05it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  3.05it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:02<00:01,  3.08it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:03<00:01,  2.97it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:03<00:00,  2.66it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:03<00:00,  2.82it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:04<00:00,  3.06it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:04<00:00,  3.09it/s]
+Evaluated 400 examples
+[2026-07-22 17:20:13,505][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.61it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.64it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.62it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.68it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.70it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.62it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.59it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.53it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.52it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.54it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.51it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.54it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.78it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.63it/s]
+Evaluated 400 examples
+[2026-07-22 17:20:16,564][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:12,  1.07s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:11,  1.07s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:10,  1.08s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:09,  1.07s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:08,  1.07s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.08s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.09s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.10s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:09<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:10<00:03,  1.10s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:12<00:02,  1.11s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.11s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.05s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.08s/it]
+Evaluated 400 examples
+[2026-07-22 17:20:30,617][metrics][INFO] - Evaluating retain_Truth_Ratio
+[2026-07-22 17:20:31,001][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.23it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.61it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  7.62it/s]
+Evaluated 100 examples
+[2026-07-22 17:20:31,764][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.53it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.52it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.11it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.52it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.55it/s]
+Evaluated 100 examples
+[2026-07-22 17:20:33,331][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+[2026-07-22 17:20:33,568][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  6.12it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  6.38it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  5.89it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  6.74it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  6.49it/s]
+Evaluated 100 examples
+[2026-07-22 17:20:34,185][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-22 17:20:34,185][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:20:34,185][metrics][INFO] - Evaluating ra_Truth_Ratio
+[2026-07-22 17:20:34,454][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.03it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.56it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  8.29it/s]
+Evaluated 117 examples
+[2026-07-22 17:20:35,184][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.56it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.49it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.09it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.56it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.57it/s]
+Evaluated 117 examples
+[2026-07-22 17:20:36,738][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+[2026-07-22 17:20:36,984][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  3.98it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  4.13it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  4.01it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  3.95it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  3.98it/s]
+Evaluated 117 examples
+[2026-07-22 17:20:37,988][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-22 17:20:37,988][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:20:37,989][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-22 17:20:37,989][metrics][INFO] - Evaluating model_utility
+[2026-07-22 17:20:37,989][evaluator][INFO] - Result for metric model_utility:	0.5306669591063944
+[2026-07-22 17:20:38,496][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:20:38,503][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.85it/s]
 31%|███       | 4/13 [00:00<00:00, 12.38it/s]
 46%|████▌     | 6/13 [00:00<00:00, 12.86it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.19it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.52it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.23it/s]
100%|██████████| 13/13 [00:00<00:00, 13.72it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 12.97it/s]
 31%|███       | 4/13 [00:00<00:00, 12.79it/s]
 46%|████▌     | 6/13 [00:00<00:00, 13.17it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.20it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.26it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.33it/s]
100%|██████████| 13/13 [00:00<00:00, 13.66it/s]
+[2026-07-22 17:20:40,406][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:20:40,413][metrics][INFO] - Evaluating privleak
+[2026-07-22 17:20:40,413][evaluator][INFO] - Result for metric privleak:	-98.06948632778128
+[2026-07-22 17:20:40,679][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:00, 11.98it/s]
Calculating ES:  31%|███       | 4/13 [00:00<00:00, 10.89it/s]
Calculating ES:  46%|████▌     | 6/13 [00:00<00:00, 11.62it/s]
Calculating ES:  62%|██████▏   | 8/13 [00:00<00:00, 12.00it/s]
Calculating ES:  77%|███████▋  | 10/13 [00:00<00:00, 12.16it/s]
Calculating ES:  92%|█████████▏| 12/13 [00:01<00:00, 11.86it/s]
Calculating ES: 100%|██████████| 13/13 [00:01<00:00, 12.31it/s]
+Evaluated 400 examples
+[2026-07-22 17:20:41,736][evaluator][INFO] - Result for metric extraction_strength:	0.32577915312947625
diff --git a/logs/eval_IdkNLL_lr5e-05_alpha10_epoch10.log b/logs/eval_IdkNLL_lr5e-05_alpha10_epoch10.log
new file mode 100644
index 0000000000000000000000000000000000000000..fc4e7d9d60d8afc8336cbf36d2bf7f606dd6137e
--- /dev/null
+++ b/logs/eval_IdkNLL_lr5e-05_alpha10_epoch10.log
@@ -0,0 +1,78 @@
+/opt/conda/envs/copygen/lib/python3.12/site-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.6.3) or chardet (6.0.0.post1)/charset_normalizer (3.4.4) doesn't match a supported version!
+  warnings.warn(
+[2026-07-22 17:14:14,089][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/IdkNLL_lr5e-05_alpha10_epoch10
+[2026-07-22 17:14:14,090][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-22 17:14:14,090][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/IdkNLL_lr5e-05_alpha10_epoch10/TOFU_EVAL.json
+[2026-07-22 17:14:14,090][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/IdkNLL_lr5e-05_alpha10_epoch10/TOFU_SUMMARY.json
+[2026-07-22 17:14:15,996][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:05,  2.24it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:03,  3.13it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  3.54it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.81it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.02it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.14it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.23it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:02<00:01,  4.36it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.40it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.44it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.42it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.40it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.60it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.15it/s]
+Evaluated 400 examples
+[2026-07-22 17:14:19,560][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:13,  1.13s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:12,  1.12s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:11,  1.15s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:10,  1.15s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:09,  1.14s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.13s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.12s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:09<00:05,  1.11s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:10<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:11<00:03,  1.10s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:12<00:02,  1.11s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.13s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.07s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.11s/it]
+Evaluated 400 examples
+[2026-07-22 17:14:33,989][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-22 17:14:33,990][evaluator][INFO] - Result for metric forget_truth_ratio:	0.5273756074800511
+[2026-07-22 17:14:33,994][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-22 17:14:33,994][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:14:34,001][metrics][INFO] - Evaluating forget_quality
+[2026-07-22 17:14:34,002][evaluator][INFO] - Result for metric forget_quality:	3.1508505165833036e-15
+[2026-07-22 17:14:35,705][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.48it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.49it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.34it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:02,  4.38it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.41it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.45it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.41it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.49it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.53it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.51it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.46it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.44it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.68it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.51it/s]
+Evaluated 400 examples
+[2026-07-22 17:14:38,591][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.60293212890625
+[2026-07-22 17:14:38,962][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:02,  4.66it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:02,  5.28it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:02,  3.92it/s]
Calculating text similarity:  31%|███       | 4/13 [00:00<00:01,  5.07it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:01,  4.05it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:01<00:01,  4.28it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:01<00:01,  4.20it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:01<00:00,  5.08it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:02<00:00,  5.05it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:02<00:00,  4.89it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:02<00:00,  4.69it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:02<00:00,  5.21it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:02<00:00,  4.78it/s]
+Evaluated 400 examples
+[2026-07-22 17:14:41,685][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.03730176358467679
+[2026-07-22 17:14:41,973][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.65it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.69it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.70it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.72it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.72it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.65it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.63it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.61it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.65it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.64it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.58it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.63it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.85it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.70it/s]
+Evaluated 400 examples
+[2026-07-22 17:14:44,992][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:04,  2.96it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:03,  3.09it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:03,  3.25it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:02,  3.33it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:02,  3.37it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:01<00:02,  3.35it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:02<00:02,  2.89it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  2.94it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:02<00:01,  2.94it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:03<00:00,  3.04it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:03<00:00,  2.88it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:03<00:00,  3.00it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:04<00:00,  3.23it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:04<00:00,  3.11it/s]
+Evaluated 400 examples
+[2026-07-22 17:14:49,438][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.35it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.45it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.43it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.55it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.61it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.55it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.52it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.48it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.48it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.50it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.48it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.51it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.75it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.56it/s]
+Evaluated 400 examples
+[2026-07-22 17:14:52,601][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:12,  1.06s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:11,  1.07s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:10,  1.08s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:09,  1.07s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:08,  1.07s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.08s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.09s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.10s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:09<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:10<00:03,  1.10s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:11<00:02,  1.11s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.10s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.06s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.08s/it]
+Evaluated 400 examples
+[2026-07-22 17:15:06,648][metrics][INFO] - Evaluating retain_Truth_Ratio
+[2026-07-22 17:15:06,995][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.18it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.59it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  7.65it/s]
+Evaluated 100 examples
+[2026-07-22 17:15:07,776][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.48it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.51it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.10it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.52it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.56it/s]
+Evaluated 100 examples
+[2026-07-22 17:15:09,342][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+[2026-07-22 17:15:09,646][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  6.41it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  6.45it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  6.56it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  5.87it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  6.08it/s]
+Evaluated 100 examples
+[2026-07-22 17:15:10,304][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-22 17:15:10,305][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:15:10,305][metrics][INFO] - Evaluating ra_Truth_Ratio
+[2026-07-22 17:15:10,673][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.05it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.56it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  8.31it/s]
+Evaluated 117 examples
+[2026-07-22 17:15:11,397][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.66it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.47it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.06it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.53it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.55it/s]
+Evaluated 117 examples
+[2026-07-22 17:15:12,968][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+[2026-07-22 17:15:13,205][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  4.41it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  4.26it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  3.84it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  4.29it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  4.21it/s]
+Evaluated 117 examples
+[2026-07-22 17:15:14,155][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-22 17:15:14,156][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:15:14,156][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-22 17:15:14,156][metrics][INFO] - Evaluating model_utility
+[2026-07-22 17:15:14,156][evaluator][INFO] - Result for metric model_utility:	0.5491056613804656
+[2026-07-22 17:15:14,780][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:15:14,786][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.98it/s]
 31%|███       | 4/13 [00:00<00:00, 12.66it/s]
 46%|████▌     | 6/13 [00:00<00:00, 13.29it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.61it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.98it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.69it/s]
100%|██████████| 13/13 [00:00<00:00, 14.18it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.36it/s]
 31%|███       | 4/13 [00:00<00:00, 13.24it/s]
 46%|████▌     | 6/13 [00:00<00:00, 13.64it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.60it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.69it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.74it/s]
100%|██████████| 13/13 [00:00<00:00, 14.06it/s]
+[2026-07-22 17:15:16,630][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:15:16,637][metrics][INFO] - Evaluating privleak
+[2026-07-22 17:15:16,637][evaluator][INFO] - Result for metric privleak:	-97.63620901012148
+[2026-07-22 17:15:16,891][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:00, 11.91it/s]
Calculating ES:  31%|███       | 4/13 [00:00<00:00, 10.92it/s]
Calculating ES:  46%|████▌     | 6/13 [00:00<00:00, 11.68it/s]
Calculating ES:  62%|██████▏   | 8/13 [00:00<00:00, 12.13it/s]
Calculating ES:  77%|███████▋  | 10/13 [00:00<00:00, 12.25it/s]
Calculating ES:  92%|█████████▏| 12/13 [00:01<00:00, 12.01it/s]
Calculating ES: 100%|██████████| 13/13 [00:01<00:00, 12.41it/s]
+Evaluated 400 examples
+[2026-07-22 17:15:17,939][evaluator][INFO] - Result for metric extraction_strength:	0.28194443899865135
diff --git a/logs/eval_IdkNLL_lr5e-05_alpha10_epoch5.log b/logs/eval_IdkNLL_lr5e-05_alpha10_epoch5.log
new file mode 100644
index 0000000000000000000000000000000000000000..6393107a3aac3192f4a872ade02711f24c21101c
--- /dev/null
+++ b/logs/eval_IdkNLL_lr5e-05_alpha10_epoch5.log
@@ -0,0 +1,78 @@
+/opt/conda/envs/copygen/lib/python3.12/site-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.6.3) or chardet (6.0.0.post1)/charset_normalizer (3.4.4) doesn't match a supported version!
+  warnings.warn(
+[2026-07-22 17:22:37,716][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/IdkNLL_lr5e-05_alpha10_epoch5
+[2026-07-22 17:22:37,717][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-22 17:22:37,717][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/IdkNLL_lr5e-05_alpha10_epoch5/TOFU_EVAL.json
+[2026-07-22 17:22:37,717][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/IdkNLL_lr5e-05_alpha10_epoch5/TOFU_SUMMARY.json
+[2026-07-22 17:22:38,404][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:05,  2.32it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:03,  3.19it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  3.57it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.81it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:02,  3.99it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.11it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.21it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:02<00:01,  4.33it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.36it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.39it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.38it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.36it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.55it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.13it/s]
+Evaluated 400 examples
+[2026-07-22 17:22:41,911][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:13,  1.13s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:12,  1.13s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:11,  1.15s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:10,  1.14s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:09,  1.13s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.13s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.12s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.11s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:10<00:04,  1.11s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:11<00:03,  1.10s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:12<00:02,  1.11s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.11s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.06s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.11s/it]
+Evaluated 400 examples
+[2026-07-22 17:22:56,284][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-22 17:22:56,284][evaluator][INFO] - Result for metric forget_truth_ratio:	0.547315924175031
+[2026-07-22 17:22:56,288][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-22 17:22:56,289][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:22:56,296][metrics][INFO] - Evaluating forget_quality
+[2026-07-22 17:22:56,297][evaluator][INFO] - Result for metric forget_quality:	5.805666105234786e-14
+[2026-07-22 17:22:56,672][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.47it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.47it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.32it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:02,  4.36it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.39it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.43it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.42it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.51it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.55it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.53it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.47it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.42it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.65it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.50it/s]
+Evaluated 400 examples
+[2026-07-22 17:22:59,564][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.57861572265625
+[2026-07-22 17:22:59,835][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:02,  4.04it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:03,  3.46it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:02,  3.51it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:02,  4.20it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:01,  4.09it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:01<00:01,  4.62it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:01<00:01,  4.68it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:01<00:01,  4.66it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:02<00:00,  4.56it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:02<00:00,  4.00it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:02<00:00,  4.13it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:02<00:00,  4.35it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  4.22it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  4.22it/s]
+Evaluated 400 examples
+[2026-07-22 17:23:02,919][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.058624543990867
+[2026-07-22 17:23:03,203][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.67it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.73it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.68it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.71it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.72it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.66it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.65it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.62it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.65it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.64it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.58it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.62it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.84it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.70it/s]
+Evaluated 400 examples
+[2026-07-22 17:23:06,321][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:04,  2.70it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:03,  2.88it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:01<00:03,  2.94it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:02,  3.16it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:02,  3.23it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:01<00:02,  3.33it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:02<00:01,  3.34it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  3.30it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:02<00:01,  3.34it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:03<00:00,  3.29it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:03<00:00,  2.99it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:03<00:00,  3.11it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:04<00:00,  3.16it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:04<00:00,  3.17it/s]
+Evaluated 400 examples
+[2026-07-22 17:23:10,719][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.63it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.64it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.61it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.67it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.69it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.58it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.54it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.49it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.49it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.50it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.47it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.50it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.71it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.59it/s]
+Evaluated 400 examples
+[2026-07-22 17:23:13,836][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:12,  1.06s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:11,  1.07s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:10,  1.08s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:09,  1.07s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:08,  1.06s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.07s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.09s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.09s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:09<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:10<00:03,  1.10s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:11<00:02,  1.10s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.10s/it]
Calculating loss: 100%|██████████| 13/13 [00:13<00:00,  1.05s/it]
Calculating loss: 100%|██████████| 13/13 [00:13<00:00,  1.08s/it]
+Evaluated 400 examples
+[2026-07-22 17:23:27,830][metrics][INFO] - Evaluating retain_Truth_Ratio
+[2026-07-22 17:23:28,169][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.13it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.50it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  7.45it/s]
+Evaluated 100 examples
+[2026-07-22 17:23:28,979][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.03it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.44it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.07it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.43it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.47it/s]
+Evaluated 100 examples
+[2026-07-22 17:23:30,601][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+[2026-07-22 17:23:30,907][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  5.00it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  5.12it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  4.78it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  5.47it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  5.26it/s]
+Evaluated 100 examples
+[2026-07-22 17:23:31,669][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-22 17:23:31,669][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:23:31,669][metrics][INFO] - Evaluating ra_Truth_Ratio
+[2026-07-22 17:23:31,948][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.01it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.45it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  8.16it/s]
+Evaluated 117 examples
+[2026-07-22 17:23:32,697][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.56it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.45it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.08it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.55it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.56it/s]
+Evaluated 117 examples
+[2026-07-22 17:23:34,259][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+[2026-07-22 17:23:34,507][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  3.55it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  4.20it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  4.11it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  4.16it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  4.10it/s]
+Evaluated 117 examples
+[2026-07-22 17:23:35,483][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-22 17:23:35,483][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:23:35,483][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-22 17:23:35,483][metrics][INFO] - Evaluating model_utility
+[2026-07-22 17:23:35,483][evaluator][INFO] - Result for metric model_utility:	0.5337837919890431
+[2026-07-22 17:23:36,017][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:23:36,023][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 14.15it/s]
 31%|███       | 4/13 [00:00<00:00, 12.86it/s]
 46%|████▌     | 6/13 [00:00<00:00, 13.37it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.82it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 14.19it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.90it/s]
100%|██████████| 13/13 [00:00<00:00, 14.36it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.12it/s]
 31%|███       | 4/13 [00:00<00:00, 13.21it/s]
 46%|████▌     | 6/13 [00:00<00:00, 13.67it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.71it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.77it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.88it/s]
100%|██████████| 13/13 [00:00<00:00, 14.18it/s]
+[2026-07-22 17:23:37,848][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:23:37,854][metrics][INFO] - Evaluating privleak
+[2026-07-22 17:23:37,854][evaluator][INFO] - Result for metric privleak:	-97.20647484809678
+[2026-07-22 17:23:38,127][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:00, 11.80it/s]
Calculating ES:  31%|███       | 4/13 [00:00<00:00, 10.53it/s]
Calculating ES:  46%|████▌     | 6/13 [00:00<00:00, 11.12it/s]
Calculating ES:  62%|██████▏   | 8/13 [00:00<00:00, 11.52it/s]
Calculating ES:  77%|███████▋  | 10/13 [00:00<00:00, 11.50it/s]
Calculating ES:  92%|█████████▏| 12/13 [00:01<00:00, 11.30it/s]
Calculating ES: 100%|██████████| 13/13 [00:01<00:00, 11.72it/s]
+Evaluated 400 examples
+[2026-07-22 17:23:39,237][evaluator][INFO] - Result for metric extraction_strength:	0.24044956048845798
diff --git a/logs/eval_IdkNLL_lr5e-05_alpha1_epoch10.log b/logs/eval_IdkNLL_lr5e-05_alpha1_epoch10.log
new file mode 100644
index 0000000000000000000000000000000000000000..381745fd136a8d013c8073df73bdf6443babb79e
--- /dev/null
+++ b/logs/eval_IdkNLL_lr5e-05_alpha1_epoch10.log
@@ -0,0 +1,78 @@
+/opt/conda/envs/copygen/lib/python3.12/site-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.6.3) or chardet (6.0.0.post1)/charset_normalizer (3.4.4) doesn't match a supported version!
+  warnings.warn(
+[2026-07-20 14:47:52,739][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/IdkNLL_lr5e-05_alpha1_epoch10
+[2026-07-20 14:47:52,741][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-20 14:47:52,741][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/IdkNLL_lr5e-05_alpha1_epoch10/TOFU_EVAL.json
+[2026-07-20 14:47:52,741][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/IdkNLL_lr5e-05_alpha1_epoch10/TOFU_SUMMARY.json
+[2026-07-20 14:47:53,675][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:09,  1.26it/s]
Calculating loss:  15%|█▌        | 2/13 [00:01<00:05,  2.01it/s]
Calculating loss:  23%|██▎       | 3/13 [00:01<00:04,  2.40it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:03,  2.67it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:02,  2.94it/s]
Calculating loss:  46%|████▌     | 6/13 [00:02<00:02,  3.09it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:02<00:01,  3.16it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:02<00:01,  3.35it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:03<00:01,  3.38it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:03<00:00,  3.40it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:03<00:00,  3.35it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:04<00:00,  3.31it/s]
Calculating loss: 100%|██████████| 13/13 [00:04<00:00,  3.45it/s]
Calculating loss: 100%|██████████| 13/13 [00:04<00:00,  3.02it/s]
+Evaluated 400 examples
+[2026-07-20 14:47:58,571][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:13,  1.13s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:12,  1.13s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:11,  1.16s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:10,  1.16s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:09,  1.14s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.13s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.13s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:09<00:05,  1.11s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:10<00:04,  1.13s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:11<00:03,  1.13s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:12<00:02,  1.14s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.10s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.04s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.11s/it]
+Evaluated 400 examples
+[2026-07-20 14:48:12,969][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-20 14:48:12,971][evaluator][INFO] - Result for metric forget_truth_ratio:	0.5344372644932901
+[2026-07-20 14:48:12,982][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-20 14:48:12,983][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-20 14:48:13,003][metrics][INFO] - Evaluating forget_quality
+[2026-07-20 14:48:13,005][evaluator][INFO] - Result for metric forget_quality:	5.805666105234786e-14
+[2026-07-20 14:48:13,443][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.78it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.77it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.65it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.68it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.69it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.72it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.36it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.53it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.65it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.68it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.70it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.50it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.83it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.67it/s]
+Evaluated 400 examples
+[2026-07-20 14:48:16,226][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.52810302734375
+[2026-07-20 14:48:16,508][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:03,  3.70it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:02,  4.25it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:02,  4.96it/s]
Calculating text similarity:  31%|███       | 4/13 [00:00<00:01,  5.52it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:00<00:01,  5.84it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:01<00:01,  6.20it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:01<00:00,  6.45it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:01<00:00,  6.64it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:01<00:00,  6.71it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:01<00:00,  6.76it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:01<00:00,  6.78it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:01<00:00,  6.71it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:02<00:00,  7.09it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:02<00:00,  6.24it/s]
+Evaluated 400 examples
+[2026-07-20 14:48:18,593][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.007173440310293194
+[2026-07-20 14:48:18,875][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.94it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.90it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.92it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.95it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.97it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.92it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.92it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.90it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.93it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.88it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.84it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.12it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  3.89it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.54it/s]
+Evaluated 400 examples
+[2026-07-20 14:48:22,002][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:01<00:12,  1.00s/it]
Calculating text similarity:  15%|█▌        | 2/13 [00:01<00:09,  1.12it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:02<00:09,  1.05it/s]
Calculating text similarity:  31%|███       | 4/13 [00:03<00:07,  1.21it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:04<00:06,  1.27it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:05<00:06,  1.12it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:06<00:05,  1.02it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:07<00:04,  1.04it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:08<00:03,  1.06it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:09<00:02,  1.06it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:10<00:02,  1.03s/it]
Calculating text similarity:  92%|█████████▏| 12/13 [00:11<00:01,  1.11s/it]
Calculating text similarity: 100%|██████████| 13/13 [00:12<00:00,  1.01s/it]
Calculating text similarity: 100%|██████████| 13/13 [00:12<00:00,  1.04it/s]
+Evaluated 400 examples
+[2026-07-20 14:48:34,885][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:05,  2.35it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:04,  2.42it/s]
Calculating loss:  23%|██▎       | 3/13 [00:01<00:03,  2.72it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:03,  2.62it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:03,  2.58it/s]
Calculating loss:  46%|████▌     | 6/13 [00:02<00:02,  2.53it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:02<00:02,  2.50it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:03<00:01,  2.67it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:03<00:01,  2.59it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:03<00:01,  2.55it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:04<00:00,  2.50it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:04<00:00,  2.49it/s]
Calculating loss: 100%|██████████| 13/13 [00:04<00:00,  2.84it/s]
Calculating loss: 100%|██████████| 13/13 [00:04<00:00,  2.62it/s]
+Evaluated 400 examples
+[2026-07-20 14:48:40,133][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:22,  1.91s/it]
Calculating loss:  15%|█▌        | 2/13 [00:03<00:21,  1.91s/it]
Calculating loss:  23%|██▎       | 3/13 [00:05<00:19,  1.94s/it]
Calculating loss:  31%|███       | 4/13 [00:07<00:15,  1.76s/it]
Calculating loss:  38%|███▊      | 5/13 [00:09<00:14,  1.78s/it]
Calculating loss:  46%|████▌     | 6/13 [00:11<00:12,  1.83s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:12<00:10,  1.79s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:14<00:08,  1.65s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:15<00:06,  1.60s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:17<00:04,  1.55s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:18<00:03,  1.52s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:20<00:01,  1.53s/it]
Calculating loss: 100%|██████████| 13/13 [00:21<00:00,  1.44s/it]
Calculating loss: 100%|██████████| 13/13 [00:21<00:00,  1.63s/it]
+Evaluated 400 examples
+[2026-07-20 14:49:01,380][metrics][INFO] - Evaluating retain_Truth_Ratio
+[2026-07-20 14:49:01,879][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  5.76it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  4.63it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  4.53it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  4.70it/s]
+Evaluated 100 examples
+[2026-07-20 14:49:03,105][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  4.31it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:01,  1.87it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  1.59it/s]
Calculating loss: 100%|██████████| 4/4 [00:02<00:00,  1.54it/s]
Calculating loss: 100%|██████████| 4/4 [00:02<00:00,  1.66it/s]
+Evaluated 100 examples
+[2026-07-20 14:49:05,511][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+[2026-07-20 14:49:05,938][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:01,  2.40it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  2.23it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:01<00:00,  2.18it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  2.12it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  2.16it/s]
+Evaluated 100 examples
+[2026-07-20 14:49:07,786][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-20 14:49:07,787][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-20 14:49:07,787][metrics][INFO] - Evaluating ra_Truth_Ratio
+[2026-07-20 14:49:08,175][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  4.59it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  3.66it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  4.59it/s]
+Evaluated 117 examples
+[2026-07-20 14:49:09,421][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  3.43it/s]
Calculating loss:  50%|█████     | 2/4 [00:01<00:01,  1.58it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  1.54it/s]
Calculating loss: 100%|██████████| 4/4 [00:02<00:00,  1.87it/s]
Calculating loss: 100%|██████████| 4/4 [00:02<00:00,  1.83it/s]
+Evaluated 117 examples
+[2026-07-20 14:49:11,610][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+[2026-07-20 14:49:12,016][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:01,  2.00it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  2.02it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:01<00:00,  1.93it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  2.05it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  2.02it/s]
+Evaluated 117 examples
+[2026-07-20 14:49:13,997][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-20 14:49:13,997][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-20 14:49:13,997][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-20 14:49:13,997][metrics][INFO] - Evaluating model_utility
+[2026-07-20 14:49:13,998][evaluator][INFO] - Result for metric model_utility:	0.4970728934935518
+[2026-07-20 14:49:14,788][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-20 14:49:14,803][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:01,  8.36it/s]
 23%|██▎       | 3/13 [00:00<00:01,  7.01it/s]
 31%|███       | 4/13 [00:00<00:01,  6.76it/s]
 38%|███▊      | 5/13 [00:00<00:01,  6.93it/s]
 46%|████▌     | 6/13 [00:00<00:01,  6.94it/s]
 54%|█████▍    | 7/13 [00:00<00:00,  7.24it/s]
 62%|██████▏   | 8/13 [00:01<00:00,  7.32it/s]
 69%|██████▉   | 9/13 [00:01<00:00,  7.38it/s]
 77%|███████▋  | 10/13 [00:01<00:00,  7.31it/s]
 85%|████████▍ | 11/13 [00:01<00:00,  7.10it/s]
 92%|█████████▏| 12/13 [00:01<00:00,  6.93it/s]
100%|██████████| 13/13 [00:01<00:00,  7.49it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
  8%|▊         | 1/13 [00:00<00:01,  6.33it/s]
 15%|█▌        | 2/13 [00:00<00:01,  6.63it/s]
 23%|██▎       | 3/13 [00:00<00:01,  6.62it/s]
 31%|███       | 4/13 [00:00<00:01,  6.95it/s]
 38%|███▊      | 5/13 [00:00<00:01,  6.75it/s]
 46%|████▌     | 6/13 [00:00<00:00,  7.25it/s]
 54%|█████▍    | 7/13 [00:00<00:00,  7.66it/s]
 62%|██████▏   | 8/13 [00:01<00:00,  7.20it/s]
 69%|██████▉   | 9/13 [00:01<00:00,  6.96it/s]
 77%|███████▋  | 10/13 [00:01<00:00,  7.09it/s]
 85%|████████▍ | 11/13 [00:01<00:00,  7.07it/s]
 92%|█████████▏| 12/13 [00:01<00:00,  7.05it/s]
100%|██████████| 13/13 [00:01<00:00,  7.37it/s]
+[2026-07-20 14:49:18,308][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-20 14:49:18,324][metrics][INFO] - Evaluating privleak
+[2026-07-20 14:49:18,325][evaluator][INFO] - Result for metric privleak:	-96.03369033288922
+[2026-07-20 14:49:18,756][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:   8%|▊         | 1/13 [00:00<00:03,  3.95it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:03,  3.37it/s]
Calculating ES:  23%|██▎       | 3/13 [00:00<00:03,  3.22it/s]
Calculating ES:  31%|███       | 4/13 [00:01<00:02,  3.03it/s]
Calculating ES:  38%|███▊      | 5/13 [00:01<00:02,  3.40it/s]
Calculating ES:  54%|█████▍    | 7/13 [00:01<00:01,  5.20it/s]
Calculating ES:  69%|██████▉   | 9/13 [00:01<00:00,  6.73it/s]
Calculating ES:  77%|███████▋  | 10/13 [00:01<00:00,  7.26it/s]
Calculating ES:  85%|████████▍ | 11/13 [00:02<00:00,  5.55it/s]
Calculating ES:  92%|█████████▏| 12/13 [00:02<00:00,  3.88it/s]
Calculating ES: 100%|██████████| 13/13 [00:03<00:00,  3.85it/s]
Calculating ES: 100%|██████████| 13/13 [00:03<00:00,  4.31it/s]
+Evaluated 400 examples
+[2026-07-20 14:49:21,771][evaluator][INFO] - Result for metric extraction_strength:	0.2364795768799366
diff --git a/logs/eval_IdkNLL_lr5e-05_alpha1_epoch5.log b/logs/eval_IdkNLL_lr5e-05_alpha1_epoch5.log
new file mode 100644
index 0000000000000000000000000000000000000000..ffbcb081315bae1a0610518d5a4251400993ca18
--- /dev/null
+++ b/logs/eval_IdkNLL_lr5e-05_alpha1_epoch5.log
@@ -0,0 +1,78 @@
+/opt/conda/envs/copygen/lib/python3.12/site-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.6.3) or chardet (6.0.0.post1)/charset_normalizer (3.4.4) doesn't match a supported version!
+  warnings.warn(
+[2026-07-22 17:21:11,260][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/IdkNLL_lr5e-05_alpha1_epoch5
+[2026-07-22 17:21:11,261][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-22 17:21:11,261][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/IdkNLL_lr5e-05_alpha1_epoch5/TOFU_EVAL.json
+[2026-07-22 17:21:11,261][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/IdkNLL_lr5e-05_alpha1_epoch5/TOFU_SUMMARY.json
+[2026-07-22 17:21:11,958][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:05,  2.18it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:03,  3.11it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  3.54it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.81it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.03it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.16it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.26it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:02<00:01,  4.39it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.44it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.48it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.46it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.44it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.64it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.16it/s]
+Evaluated 400 examples
+[2026-07-22 17:21:15,651][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:13,  1.12s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:12,  1.12s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:11,  1.15s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:10,  1.15s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:09,  1.14s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.13s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.12s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.10s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:10<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:11<00:03,  1.10s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:12<00:02,  1.11s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.12s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.06s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.10s/it]
+Evaluated 400 examples
+[2026-07-22 17:21:30,016][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-22 17:21:30,016][evaluator][INFO] - Result for metric forget_truth_ratio:	0.5453775468072528
+[2026-07-22 17:21:30,020][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-22 17:21:30,021][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:21:30,028][metrics][INFO] - Evaluating forget_quality
+[2026-07-22 17:21:30,028][evaluator][INFO] - Result for metric forget_quality:	1.02724452984523e-14
+[2026-07-22 17:21:30,372][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.51it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.51it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.36it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:02,  4.39it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.43it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.48it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.47it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.56it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.56it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.55it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.51it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.49it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.73it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.55it/s]
+Evaluated 400 examples
+[2026-07-22 17:21:33,232][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.52435302734375
+[2026-07-22 17:21:33,483][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:01,  8.42it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:00, 10.00it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:00<00:00, 10.36it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:00<00:00, 10.90it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:00<00:00, 10.97it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:01<00:00, 10.78it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:01<00:00, 10.96it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:01<00:00, 10.73it/s]
+Evaluated 400 examples
+[2026-07-22 17:21:34,694][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.005011347532056894
+[2026-07-22 17:21:34,949][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.56it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.67it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.67it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.70it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.66it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.57it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.56it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.54it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.60it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.62it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.58it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.63it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.86it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.67it/s]
+Evaluated 400 examples
+[2026-07-22 17:21:38,069][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:03,  3.55it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:03,  3.56it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:02,  3.63it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:02,  3.83it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:02,  3.85it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:01<00:01,  3.82it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:01<00:01,  3.63it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  3.50it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:02<00:01,  3.36it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:02<00:00,  3.41it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:03<00:00,  3.28it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:03<00:00,  3.39it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.39it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.50it/s]
+Evaluated 400 examples
+[2026-07-22 17:21:42,050][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.53it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.59it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.56it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.62it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.65it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.57it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.53it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.48it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.45it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.44it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.40it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.43it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.68it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.55it/s]
+Evaluated 400 examples
+[2026-07-22 17:21:45,166][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:12,  1.06s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:11,  1.06s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:10,  1.07s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:09,  1.07s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:08,  1.06s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.07s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.09s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.10s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:09<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:10<00:03,  1.09s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:11<00:02,  1.10s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.10s/it]
Calculating loss: 100%|██████████| 13/13 [00:13<00:00,  1.05s/it]
Calculating loss: 100%|██████████| 13/13 [00:13<00:00,  1.07s/it]
+Evaluated 400 examples
+[2026-07-22 17:21:59,124][metrics][INFO] - Evaluating retain_Truth_Ratio
+[2026-07-22 17:21:59,478][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  7.99it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.45it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  7.46it/s]
+Evaluated 100 examples
+[2026-07-22 17:22:00,293][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.09it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.43it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.06it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.48it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.50it/s]
+Evaluated 100 examples
+[2026-07-22 17:22:01,894][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+[2026-07-22 17:22:02,370][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  6.17it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  5.98it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  6.19it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  7.22it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  6.77it/s]
+Evaluated 100 examples
+[2026-07-22 17:22:02,962][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-22 17:22:02,962][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:22:02,962][metrics][INFO] - Evaluating ra_Truth_Ratio
+[2026-07-22 17:22:03,262][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.13it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.62it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  8.37it/s]
+Evaluated 117 examples
+[2026-07-22 17:22:04,244][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.71it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.50it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.09it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.56it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.58it/s]
+Evaluated 117 examples
+[2026-07-22 17:22:05,797][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+[2026-07-22 17:22:06,106][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  4.01it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  4.38it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  3.40it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  3.94it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  3.89it/s]
+Evaluated 117 examples
+[2026-07-22 17:22:07,135][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-22 17:22:07,136][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:22:07,136][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-22 17:22:07,136][metrics][INFO] - Evaluating model_utility
+[2026-07-22 17:22:07,136][evaluator][INFO] - Result for metric model_utility:	0.4558601581687516
+[2026-07-22 17:22:07,675][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:22:07,686][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.16it/s]
 31%|███       | 4/13 [00:00<00:00, 12.08it/s]
 46%|████▌     | 6/13 [00:00<00:00, 12.67it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 12.85it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.16it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 12.98it/s]
100%|██████████| 13/13 [00:00<00:00, 13.34it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 12.64it/s]
 31%|███       | 4/13 [00:00<00:00, 12.31it/s]
 46%|████▌     | 6/13 [00:00<00:00, 12.64it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 12.70it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 12.80it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 12.90it/s]
100%|██████████| 13/13 [00:00<00:00, 13.21it/s]
+[2026-07-22 17:22:09,647][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:22:09,654][metrics][INFO] - Evaluating privleak
+[2026-07-22 17:22:09,654][evaluator][INFO] - Result for metric privleak:	-95.85248323040999
+[2026-07-22 17:22:10,017][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:00, 11.49it/s]
Calculating ES:  31%|███       | 4/13 [00:00<00:00, 10.63it/s]
Calculating ES:  46%|████▌     | 6/13 [00:00<00:00, 11.32it/s]
Calculating ES:  62%|██████▏   | 8/13 [00:00<00:00, 11.73it/s]
Calculating ES:  77%|███████▋  | 10/13 [00:00<00:00, 11.90it/s]
Calculating ES:  92%|█████████▏| 12/13 [00:01<00:00, 11.68it/s]
Calculating ES: 100%|██████████| 13/13 [00:01<00:00, 12.06it/s]
+Evaluated 400 examples
+[2026-07-22 17:22:11,096][evaluator][INFO] - Result for metric extraction_strength:	0.2176950903433096
diff --git a/logs/eval_IdkNLL_lr5e-05_alpha2_epoch10.log b/logs/eval_IdkNLL_lr5e-05_alpha2_epoch10.log
new file mode 100644
index 0000000000000000000000000000000000000000..2dac31e9ab1c42eb5b6f0c8a78bfa2b23f1bc272
--- /dev/null
+++ b/logs/eval_IdkNLL_lr5e-05_alpha2_epoch10.log
@@ -0,0 +1,76 @@
+[2026-07-21 07:50:33,476][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/IdkNLL_lr5e-05_alpha2_epoch10
+[2026-07-21 07:50:33,477][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-21 07:50:33,477][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/IdkNLL_lr5e-05_alpha2_epoch10/TOFU_EVAL.json
+[2026-07-21 07:50:33,477][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/IdkNLL_lr5e-05_alpha2_epoch10/TOFU_SUMMARY.json
+[2026-07-21 07:50:34,111][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:07,  1.51it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:04,  2.51it/s]
Calculating loss:  23%|██▎       | 3/13 [00:01<00:03,  3.15it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.60it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:02,  3.95it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.14it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.33it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:02<00:01,  4.50it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.57it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.62it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.63it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:03<00:00,  4.63it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.80it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.06it/s]
+Evaluated 400 examples
+[2026-07-21 07:50:37,648][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:12,  1.03s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:11,  1.03s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:10,  1.05s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:09,  1.05s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:08,  1.04s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.03s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.03s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.02s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:09<00:04,  1.02s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:10<00:03,  1.02s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:11<00:02,  1.02s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:12<00:01,  1.03s/it]
Calculating loss: 100%|██████████| 13/13 [00:13<00:00,  1.03it/s]
Calculating loss: 100%|██████████| 13/13 [00:13<00:00,  1.02s/it]
+Evaluated 400 examples
+[2026-07-21 07:50:50,850][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-21 07:50:50,851][evaluator][INFO] - Result for metric forget_truth_ratio:	0.5321104903862071
+[2026-07-21 07:50:50,859][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-21 07:50:50,860][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-21 07:50:50,875][metrics][INFO] - Evaluating forget_quality
+[2026-07-21 07:50:50,877][evaluator][INFO] - Result for metric forget_quality:	5.7046977773929285e-15
+[2026-07-21 07:50:51,263][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.78it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.78it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.65it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.68it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.70it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.74it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.74it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.82it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.85it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.83it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.81it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.79it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  5.05it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.84it/s]
+Evaluated 400 examples
+[2026-07-21 07:50:53,950][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.53903564453125
+[2026-07-21 07:50:54,250][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:04,  2.96it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:02,  4.44it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:01,  5.10it/s]
Calculating text similarity:  31%|███       | 4/13 [00:00<00:01,  5.67it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:00<00:01,  5.97it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:01<00:01,  6.31it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:01<00:00,  6.55it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:01<00:00,  6.73it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:01<00:00,  4.76it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:02<00:00,  4.06it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:02<00:00,  3.53it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:02<00:00,  4.10it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:02<00:00,  4.85it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:02<00:00,  4.85it/s]
+Evaluated 400 examples
+[2026-07-21 07:50:56,929][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.014969829593524582
+[2026-07-21 07:50:57,323][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.94it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  5.00it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:01,  5.01it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  5.01it/s]
Calculating loss:  38%|███▊      | 5/13 [00:00<00:01,  5.01it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.94it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.93it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.90it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.93it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.92it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.87it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.91it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  5.15it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.99it/s]
+Evaluated 400 examples
+[2026-07-21 07:51:00,223][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:07,  1.59it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:01<00:06,  1.63it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:01<00:05,  1.73it/s]
Calculating text similarity:  31%|███       | 4/13 [00:02<00:05,  1.77it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:02<00:04,  1.77it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:03<00:04,  1.65it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:04<00:03,  1.60it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:04<00:03,  1.63it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:05<00:02,  1.72it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:05<00:01,  1.71it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:06<00:01,  1.61it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:07<00:00,  1.65it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:07<00:00,  1.78it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:07<00:00,  1.70it/s]
+Evaluated 400 examples
+[2026-07-21 07:51:08,233][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.87it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.89it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.87it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.92it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.95it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.87it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.85it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.80it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.79it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.80it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.78it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.80it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  5.06it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.89it/s]
+Evaluated 400 examples
+[2026-07-21 07:51:11,167][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:11,  1.01it/s]
Calculating loss:  15%|█▌        | 2/13 [00:01<00:10,  1.01it/s]
Calculating loss:  23%|██▎       | 3/13 [00:02<00:09,  1.00it/s]
Calculating loss:  31%|███       | 4/13 [00:03<00:08,  1.01it/s]
Calculating loss:  38%|███▊      | 5/13 [00:04<00:07,  1.02it/s]
Calculating loss:  46%|████▌     | 6/13 [00:05<00:06,  1.01it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:06<00:06,  1.00s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:07<00:05,  1.01s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:09<00:04,  1.01s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:10<00:03,  1.01s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:11<00:02,  1.02s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:12<00:01,  1.01s/it]
Calculating loss: 100%|██████████| 13/13 [00:12<00:00,  1.03it/s]
Calculating loss: 100%|██████████| 13/13 [00:12<00:00,  1.01it/s]
+Evaluated 400 examples
+[2026-07-21 07:51:24,076][metrics][INFO] - Evaluating retain_Truth_Ratio
+[2026-07-21 07:51:24,446][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.37it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.81it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  7.95it/s]
+Evaluated 100 examples
+[2026-07-21 07:51:25,259][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.28it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.61it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.21it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.63it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.66it/s]
+Evaluated 100 examples
+[2026-07-21 07:51:26,765][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+[2026-07-21 07:51:27,027][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:01,  2.61it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  2.86it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:01<00:00,  2.84it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  3.06it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  2.96it/s]
+Evaluated 100 examples
+[2026-07-21 07:51:28,381][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-21 07:51:28,381][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-21 07:51:28,382][metrics][INFO] - Evaluating ra_Truth_Ratio
+[2026-07-21 07:51:28,642][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.31it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.84it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  8.58it/s]
+Evaluated 117 examples
+[2026-07-21 07:51:29,382][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.43it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.60it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.21it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.70it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.70it/s]
+Evaluated 117 examples
+[2026-07-21 07:51:30,863][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+[2026-07-21 07:51:31,124][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:01,  2.31it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  2.40it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:01<00:00,  2.31it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  2.34it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  2.34it/s]
+Evaluated 117 examples
+[2026-07-21 07:51:32,832][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-21 07:51:32,833][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-21 07:51:32,833][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-21 07:51:32,833][metrics][INFO] - Evaluating model_utility
+[2026-07-21 07:51:32,834][evaluator][INFO] - Result for metric model_utility:	0.5362012798873454
+[2026-07-21 07:51:33,451][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-21 07:51:33,466][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 12.94it/s]
 31%|███       | 4/13 [00:00<00:00, 11.05it/s]
 46%|████▌     | 6/13 [00:00<00:00, 11.31it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 11.44it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 11.70it/s]
 92%|█████████▏| 12/13 [00:01<00:00, 11.44it/s]
100%|██████████| 13/13 [00:01<00:00, 11.97it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 11.21it/s]
 31%|███       | 4/13 [00:00<00:00, 11.06it/s]
 46%|████▌     | 6/13 [00:00<00:00, 11.36it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 11.42it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 11.42it/s]
 92%|█████████▏| 12/13 [00:01<00:00, 11.54it/s]
100%|██████████| 13/13 [00:01<00:00, 11.82it/s]
+[2026-07-21 07:51:35,657][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-21 07:51:35,671][metrics][INFO] - Evaluating privleak
+[2026-07-21 07:51:35,671][evaluator][INFO] - Result for metric privleak:	-96.29993317061005
+[2026-07-21 07:51:35,984][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:01, 10.45it/s]
Calculating ES:  31%|███       | 4/13 [00:00<00:00,  9.34it/s]
Calculating ES:  46%|████▌     | 6/13 [00:00<00:00,  9.81it/s]
Calculating ES:  54%|█████▍    | 7/13 [00:00<00:00,  9.81it/s]
Calculating ES:  69%|██████▉   | 9/13 [00:00<00:00, 10.30it/s]
Calculating ES:  85%|████████▍ | 11/13 [00:01<00:00, 10.06it/s]
Calculating ES: 100%|██████████| 13/13 [00:01<00:00, 10.82it/s]
Calculating ES: 100%|██████████| 13/13 [00:01<00:00, 10.31it/s]
+Evaluated 400 examples
+[2026-07-21 07:51:37,245][evaluator][INFO] - Result for metric extraction_strength:	0.2512007516134813
diff --git a/logs/eval_IdkNLL_lr5e-05_alpha2_epoch5.log b/logs/eval_IdkNLL_lr5e-05_alpha2_epoch5.log
new file mode 100644
index 0000000000000000000000000000000000000000..9c3fe58673cd0a083a7de6ebcb01f4b9ef333d7c
--- /dev/null
+++ b/logs/eval_IdkNLL_lr5e-05_alpha2_epoch5.log
@@ -0,0 +1,78 @@
+/opt/conda/envs/copygen/lib/python3.12/site-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.6.3) or chardet (6.0.0.post1)/charset_normalizer (3.4.4) doesn't match a supported version!
+  warnings.warn(
+[2026-07-22 17:15:45,417][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/IdkNLL_lr5e-05_alpha2_epoch5
+[2026-07-22 17:15:45,418][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-22 17:15:45,418][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/IdkNLL_lr5e-05_alpha2_epoch5/TOFU_EVAL.json
+[2026-07-22 17:15:45,418][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/IdkNLL_lr5e-05_alpha2_epoch5/TOFU_SUMMARY.json
+[2026-07-22 17:15:46,065][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:05,  2.31it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:03,  3.20it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  3.56it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.77it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:02,  3.96it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.06it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.16it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:02<00:01,  4.30it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.37it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.42it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.41it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.40it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.61it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.13it/s]
+Evaluated 400 examples
+[2026-07-22 17:15:49,550][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:13,  1.11s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:12,  1.12s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:11,  1.15s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:10,  1.15s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:09,  1.13s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.14s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.13s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.11s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:10<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:11<00:03,  1.11s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:12<00:02,  1.12s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.13s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.07s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.11s/it]
+Evaluated 400 examples
+[2026-07-22 17:16:03,998][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-22 17:16:03,999][evaluator][INFO] - Result for metric forget_truth_ratio:	0.5486427139963626
+[2026-07-22 17:16:04,003][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-22 17:16:04,004][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:16:04,012][metrics][INFO] - Evaluating forget_quality
+[2026-07-22 17:16:04,013][evaluator][INFO] - Result for metric forget_quality:	1.0229984068189572e-13
+[2026-07-22 17:16:04,466][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.47it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.47it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.31it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:02,  4.35it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.39it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.44it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.43it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.53it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.57it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.56it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.53it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.50it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.74it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.53it/s]
+Evaluated 400 examples
+[2026-07-22 17:16:07,333][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.52466552734375
+[2026-07-22 17:16:08,272][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:02,  5.13it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:01,  8.61it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:00<00:00,  9.49it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:00<00:00,  9.78it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:00<00:00,  9.78it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:01<00:00,  7.58it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:01<00:00,  8.38it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:01<00:00,  8.80it/s]
+Evaluated 400 examples
+[2026-07-22 17:16:09,749][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.006607161672815781
+[2026-07-22 17:16:10,409][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.56it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.63it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.66it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.68it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.69it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.63it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.56it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.51it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.52it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.50it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.47it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.53it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.76it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.62it/s]
+Evaluated 400 examples
+[2026-07-22 17:16:13,537][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:03,  3.69it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:03,  3.45it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:02,  3.48it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:02,  3.61it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:02,  3.53it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:01<00:02,  3.40it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:02<00:01,  3.34it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  3.16it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:02<00:01,  3.31it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:03<00:01,  2.89it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:03<00:00,  2.89it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:03<00:00,  2.96it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:04<00:00,  3.15it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:04<00:00,  3.22it/s]
+Evaluated 400 examples
+[2026-07-22 17:16:18,303][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.50it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.55it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.53it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.60it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.64it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.57it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.53it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.47it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.47it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.48it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.40it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.41it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.64it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.54it/s]
+Evaluated 400 examples
+[2026-07-22 17:16:22,394][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:12,  1.06s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:11,  1.06s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:10,  1.08s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:09,  1.07s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:08,  1.06s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.08s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.09s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.10s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:09<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:10<00:03,  1.09s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:11<00:02,  1.10s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.10s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.05s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.08s/it]
+Evaluated 400 examples
+[2026-07-22 17:16:36,399][metrics][INFO] - Evaluating retain_Truth_Ratio
+[2026-07-22 17:16:36,751][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  7.79it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.32it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  6.33it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  6.51it/s]
+Evaluated 100 examples
+[2026-07-22 17:16:37,615][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.45it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.51it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.10it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.52it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.55it/s]
+Evaluated 100 examples
+[2026-07-22 17:16:39,184][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+[2026-07-22 17:16:39,707][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  6.62it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  5.60it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  5.00it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  5.89it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  5.73it/s]
+Evaluated 100 examples
+[2026-07-22 17:16:40,405][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-22 17:16:40,405][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:16:40,405][metrics][INFO] - Evaluating ra_Truth_Ratio
+[2026-07-22 17:16:40,818][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.06it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.53it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  8.30it/s]
+Evaluated 117 examples
+[2026-07-22 17:16:41,572][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.36it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.45it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.07it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.54it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.55it/s]
+Evaluated 117 examples
+[2026-07-22 17:16:43,142][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+[2026-07-22 17:16:43,411][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:01,  2.94it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  3.61it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  3.87it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  4.01it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  3.83it/s]
+Evaluated 117 examples
+[2026-07-22 17:16:44,456][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-22 17:16:44,456][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:16:44,456][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-22 17:16:44,456][metrics][INFO] - Evaluating model_utility
+[2026-07-22 17:16:44,457][evaluator][INFO] - Result for metric model_utility:	0.5057544347432658
+[2026-07-22 17:16:44,993][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:16:45,002][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.47it/s]
 31%|███       | 4/13 [00:00<00:00, 12.17it/s]
 46%|████▌     | 6/13 [00:00<00:00, 12.77it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.24it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.70it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.49it/s]
100%|██████████| 13/13 [00:00<00:00, 13.87it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.39it/s]
 31%|███       | 4/13 [00:00<00:00, 13.22it/s]
 46%|████▌     | 6/13 [00:00<00:00, 13.64it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.66it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.72it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.81it/s]
100%|██████████| 13/13 [00:00<00:00, 14.15it/s]
+[2026-07-22 17:16:46,860][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:16:46,867][metrics][INFO] - Evaluating privleak
+[2026-07-22 17:16:46,867][evaluator][INFO] - Result for metric privleak:	-95.55688281742715
+[2026-07-22 17:17:12,006][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:00, 11.60it/s]
Calculating ES:  31%|███       | 4/13 [00:00<00:00, 10.81it/s]
Calculating ES:  46%|████▌     | 6/13 [00:00<00:00, 11.60it/s]
Calculating ES:  62%|██████▏   | 8/13 [00:00<00:00, 12.09it/s]
Calculating ES:  77%|███████▋  | 10/13 [00:00<00:00, 12.26it/s]
Calculating ES:  92%|█████████▏| 12/13 [00:01<00:00, 12.04it/s]
Calculating ES: 100%|██████████| 13/13 [00:01<00:00, 12.39it/s]
+Evaluated 400 examples
+[2026-07-22 17:17:13,056][evaluator][INFO] - Result for metric extraction_strength:	0.21479457793325007
diff --git a/logs/eval_IdkNLL_lr5e-05_alpha5_epoch10.log b/logs/eval_IdkNLL_lr5e-05_alpha5_epoch10.log
new file mode 100644
index 0000000000000000000000000000000000000000..1b85dd413959330a577651272137e3277522af54
--- /dev/null
+++ b/logs/eval_IdkNLL_lr5e-05_alpha5_epoch10.log
@@ -0,0 +1,78 @@
+/opt/conda/envs/copygen/lib/python3.12/site-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.6.3) or chardet (6.0.0.post1)/charset_normalizer (3.4.4) doesn't match a supported version!
+  warnings.warn(
+[2026-07-20 14:51:46,440][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/IdkNLL_lr5e-05_alpha5_epoch10
+[2026-07-20 14:51:46,441][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-20 14:51:46,441][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/IdkNLL_lr5e-05_alpha5_epoch10/TOFU_EVAL.json
+[2026-07-20 14:51:46,441][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/IdkNLL_lr5e-05_alpha5_epoch10/TOFU_SUMMARY.json
+[2026-07-20 14:51:47,120][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:09,  1.20it/s]
Calculating loss:  15%|█▌        | 2/13 [00:01<00:06,  1.70it/s]
Calculating loss:  23%|██▎       | 3/13 [00:01<00:05,  1.92it/s]
Calculating loss:  31%|███       | 4/13 [00:02<00:04,  2.02it/s]
Calculating loss:  38%|███▊      | 5/13 [00:02<00:03,  2.14it/s]
Calculating loss:  46%|████▌     | 6/13 [00:02<00:03,  2.27it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:03<00:02,  2.30it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:03<00:02,  2.24it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:04<00:01,  2.28it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:04<00:01,  2.50it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:05<00:00,  2.43it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:05<00:00,  2.41it/s]
Calculating loss: 100%|██████████| 13/13 [00:05<00:00,  2.49it/s]
Calculating loss: 100%|██████████| 13/13 [00:05<00:00,  2.24it/s]
+Evaluated 400 examples
+[2026-07-20 14:51:53,582][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:22,  1.88s/it]
Calculating loss:  15%|█▌        | 2/13 [00:03<00:21,  1.99s/it]
Calculating loss:  23%|██▎       | 3/13 [00:06<00:20,  2.05s/it]
Calculating loss:  31%|███       | 4/13 [00:08<00:18,  2.04s/it]
Calculating loss:  38%|███▊      | 5/13 [00:10<00:16,  2.06s/it]
Calculating loss:  46%|████▌     | 6/13 [00:12<00:14,  2.05s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:14<00:12,  2.09s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:15<00:09,  1.90s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:17<00:07,  1.81s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:19<00:05,  1.86s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:20<00:03,  1.73s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:22<00:01,  1.66s/it]
Calculating loss: 100%|██████████| 13/13 [00:23<00:00,  1.55s/it]
Calculating loss: 100%|██████████| 13/13 [00:23<00:00,  1.82s/it]
+Evaluated 400 examples
+[2026-07-20 14:52:17,272][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-20 14:52:17,274][evaluator][INFO] - Result for metric forget_truth_ratio:	0.5298605825153447
+[2026-07-20 14:52:17,285][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-20 14:52:17,286][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-20 14:52:17,306][metrics][INFO] - Evaluating forget_quality
+[2026-07-20 14:52:17,308][evaluator][INFO] - Result for metric forget_quality:	5.7046977773929285e-15
+[2026-07-20 14:52:18,007][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:03,  3.69it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:03,  3.56it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:03,  3.25it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.17it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:02,  3.14it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:02,  3.13it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:02<00:01,  3.10it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:02<00:01,  3.13it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:01,  3.13it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:03<00:00,  3.31it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:03<00:00,  3.41it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:03<00:00,  3.42it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  3.77it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  3.37it/s]
+Evaluated 400 examples
+[2026-07-20 14:52:21,870][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.56814697265625
+[2026-07-20 14:52:22,414][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:04,  2.41it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:04,  2.58it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:01<00:03,  2.50it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:03,  2.50it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:02<00:04,  1.82it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:02<00:02,  2.41it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:02<00:01,  3.06it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  3.72it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:03<00:01,  2.98it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:03<00:01,  2.64it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:04<00:00,  2.17it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:04<00:00,  2.40it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:05<00:00,  2.61it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:05<00:00,  2.56it/s]
+Evaluated 400 examples
+[2026-07-20 14:52:27,498][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.017253889918628383
+[2026-07-20 14:52:28,032][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:04,  2.81it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:03,  3.17it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  3.81it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  4.18it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.33it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  3.90it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  3.86it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:02<00:01,  3.91it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:01,  3.72it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  3.39it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:03<00:00,  3.31it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:03<00:00,  3.38it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  3.68it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  3.66it/s]
+Evaluated 400 examples
+[2026-07-20 14:52:31,912][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:11,  1.06it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:01<00:10,  1.06it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:02<00:08,  1.20it/s]
Calculating text similarity:  31%|███       | 4/13 [00:03<00:07,  1.28it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:04<00:07,  1.08it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:05<00:06,  1.12it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:06<00:05,  1.15it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:07<00:04,  1.10it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:08<00:04,  1.03s/it]
Calculating text similarity:  77%|███████▋  | 10/13 [00:09<00:02,  1.01it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:10<00:02,  1.05s/it]
Calculating text similarity:  92%|█████████▏| 12/13 [00:11<00:01,  1.08s/it]
Calculating text similarity: 100%|██████████| 13/13 [00:12<00:00,  1.04s/it]
Calculating text similarity: 100%|██████████| 13/13 [00:12<00:00,  1.03it/s]
+Evaluated 400 examples
+[2026-07-20 14:52:45,111][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:03,  3.72it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:03,  3.38it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  3.52it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.62it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:02,  3.66it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  3.60it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  3.84it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:02<00:01,  3.66it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:01,  3.63it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  3.86it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  3.88it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:03<00:00,  3.74it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  3.56it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  3.66it/s]
+Evaluated 400 examples
+[2026-07-20 14:52:49,024][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:22,  1.89s/it]
Calculating loss:  15%|█▌        | 2/13 [00:03<00:21,  1.92s/it]
Calculating loss:  23%|██▎       | 3/13 [00:05<00:18,  1.83s/it]
Calculating loss:  31%|███       | 4/13 [00:07<00:15,  1.71s/it]
Calculating loss:  38%|███▊      | 5/13 [00:08<00:14,  1.77s/it]
Calculating loss:  46%|████▌     | 6/13 [00:10<00:12,  1.81s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:12<00:10,  1.77s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:14<00:09,  1.84s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:16<00:07,  1.88s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:18<00:05,  1.95s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:20<00:03,  1.98s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:22<00:01,  1.97s/it]
Calculating loss: 100%|██████████| 13/13 [00:24<00:00,  1.89s/it]
Calculating loss: 100%|██████████| 13/13 [00:24<00:00,  1.87s/it]
+Evaluated 400 examples
+[2026-07-20 14:53:13,333][metrics][INFO] - Evaluating retain_Truth_Ratio
+[2026-07-20 14:53:13,726][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  4.14it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  3.01it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  3.13it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  3.22it/s]
+Evaluated 100 examples
+[2026-07-20 14:53:15,265][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  3.21it/s]
Calculating loss:  50%|█████     | 2/4 [00:01<00:01,  1.30it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:02<00:00,  1.16it/s]
Calculating loss: 100%|██████████| 4/4 [00:03<00:00,  1.11it/s]
Calculating loss: 100%|██████████| 4/4 [00:03<00:00,  1.20it/s]
+Evaluated 100 examples
+[2026-07-20 14:53:18,598][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+[2026-07-20 14:53:19,078][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:02,  1.48it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:01<00:01,  1.06it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:02<00:00,  1.20it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:03<00:00,  1.40it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:03<00:00,  1.31it/s]
+Evaluated 100 examples
+[2026-07-20 14:53:22,122][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-20 14:53:22,122][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-20 14:53:22,122][metrics][INFO] - Evaluating ra_Truth_Ratio
+[2026-07-20 14:53:22,584][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  9.14it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  3.67it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  4.07it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  4.95it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  4.73it/s]
+Evaluated 117 examples
+[2026-07-20 14:53:23,694][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  3.21it/s]
Calculating loss:  50%|█████     | 2/4 [00:01<00:01,  1.76it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  1.42it/s]
Calculating loss: 100%|██████████| 4/4 [00:02<00:00,  1.56it/s]
Calculating loss: 100%|██████████| 4/4 [00:02<00:00,  1.62it/s]
+Evaluated 117 examples
+[2026-07-20 14:53:26,169][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+[2026-07-20 14:53:26,503][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:01<00:03,  1.05s/it]
Calculating text similarity:  50%|█████     | 2/4 [00:01<00:01,  1.15it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:02<00:00,  1.08it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:03<00:00,  1.32it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:03<00:00,  1.22it/s]
+Evaluated 117 examples
+[2026-07-20 14:53:29,788][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-20 14:53:29,789][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-20 14:53:29,789][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-20 14:53:29,789][metrics][INFO] - Evaluating model_utility
+[2026-07-20 14:53:29,790][evaluator][INFO] - Result for metric model_utility:	0.5547274672864976
+[2026-07-20 14:53:30,439][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-20 14:53:30,455][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
  8%|▊         | 1/13 [00:00<00:01,  8.11it/s]
 15%|█▌        | 2/13 [00:00<00:01,  7.43it/s]
 23%|██▎       | 3/13 [00:00<00:01,  6.73it/s]
 31%|███       | 4/13 [00:00<00:01,  5.91it/s]
 38%|███▊      | 5/13 [00:00<00:01,  6.08it/s]
 46%|████▌     | 6/13 [00:00<00:01,  6.33it/s]
 54%|█████▍    | 7/13 [00:01<00:01,  5.86it/s]
 62%|██████▏   | 8/13 [00:01<00:00,  6.08it/s]
 69%|██████▉   | 9/13 [00:01<00:00,  6.35it/s]
 77%|███████▋  | 10/13 [00:01<00:00,  6.28it/s]
 85%|████████▍ | 11/13 [00:01<00:00,  6.11it/s]
100%|██████████| 13/13 [00:01<00:00,  8.39it/s]
100%|██████████| 13/13 [00:01<00:00,  6.85it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
  8%|▊         | 1/13 [00:00<00:01,  6.09it/s]
 15%|█▌        | 2/13 [00:00<00:01,  6.87it/s]
 23%|██▎       | 3/13 [00:00<00:01,  7.10it/s]
 31%|███       | 4/13 [00:00<00:01,  6.11it/s]
 38%|███▊      | 5/13 [00:00<00:01,  5.69it/s]
 46%|████▌     | 6/13 [00:00<00:01,  6.58it/s]
 54%|█████▍    | 7/13 [00:01<00:00,  6.97it/s]
 62%|██████▏   | 8/13 [00:01<00:00,  6.89it/s]
 69%|██████▉   | 9/13 [00:01<00:00,  6.67it/s]
 77%|███████▋  | 10/13 [00:01<00:00,  6.77it/s]
 85%|████████▍ | 11/13 [00:01<00:00,  6.71it/s]
 92%|█████████▏| 12/13 [00:01<00:00,  6.59it/s]
100%|██████████| 13/13 [00:01<00:00,  6.81it/s]
+[2026-07-20 14:53:34,269][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-20 14:53:34,287][metrics][INFO] - Evaluating privleak
+[2026-07-20 14:53:34,287][evaluator][INFO] - Result for metric privleak:	-96.96604643000288
+[2026-07-20 14:53:34,581][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:   8%|▊         | 1/13 [00:00<00:04,  2.53it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:04,  2.44it/s]
Calculating ES:  23%|██▎       | 3/13 [00:01<00:04,  2.42it/s]
Calculating ES:  31%|███       | 4/13 [00:01<00:02,  3.43it/s]
Calculating ES:  38%|███▊      | 5/13 [00:01<00:02,  3.23it/s]
Calculating ES:  46%|████▌     | 6/13 [00:01<00:02,  3.31it/s]
Calculating ES:  54%|█████▍    | 7/13 [00:02<00:01,  3.12it/s]
Calculating ES:  62%|██████▏   | 8/13 [00:02<00:01,  3.27it/s]
Calculating ES:  69%|██████▉   | 9/13 [00:02<00:01,  3.10it/s]
Calculating ES:  77%|███████▋  | 10/13 [00:03<00:01,  2.59it/s]
Calculating ES:  85%|████████▍ | 11/13 [00:03<00:00,  2.55it/s]
Calculating ES:  92%|█████████▏| 12/13 [00:04<00:00,  2.98it/s]
Calculating ES: 100%|██████████| 13/13 [00:04<00:00,  3.10it/s]
+Evaluated 400 examples
+[2026-07-20 14:53:38,773][evaluator][INFO] - Result for metric extraction_strength:	0.25835151129389794
diff --git a/logs/eval_IdkNLL_lr5e-05_alpha5_epoch5.log b/logs/eval_IdkNLL_lr5e-05_alpha5_epoch5.log
new file mode 100644
index 0000000000000000000000000000000000000000..a048486a624f78f464f513427fa3fbddf5e83f16
--- /dev/null
+++ b/logs/eval_IdkNLL_lr5e-05_alpha5_epoch5.log
@@ -0,0 +1,78 @@
+/opt/conda/envs/copygen/lib/python3.12/site-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.6.3) or chardet (6.0.0.post1)/charset_normalizer (3.4.4) doesn't match a supported version!
+  warnings.warn(
+[2026-07-22 17:24:07,257][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/IdkNLL_lr5e-05_alpha5_epoch5
+[2026-07-22 17:24:07,258][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-22 17:24:07,258][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/IdkNLL_lr5e-05_alpha5_epoch5/TOFU_EVAL.json
+[2026-07-22 17:24:07,258][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/IdkNLL_lr5e-05_alpha5_epoch5/TOFU_SUMMARY.json
+[2026-07-22 17:24:08,026][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:04,  2.48it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:03,  3.34it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  3.69it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.91it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.09it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.19it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.28it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.38it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.41it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.45it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.43it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.41it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.61it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.21it/s]
+Evaluated 400 examples
+[2026-07-22 17:24:11,521][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:13,  1.12s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:12,  1.12s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:11,  1.15s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:10,  1.15s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:09,  1.14s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.13s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.12s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.11s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:10<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:11<00:03,  1.10s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:12<00:02,  1.12s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.12s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.06s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.11s/it]
+Evaluated 400 examples
+[2026-07-22 17:24:25,908][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-22 17:24:25,909][evaluator][INFO] - Result for metric forget_truth_ratio:	0.5429714470601403
+[2026-07-22 17:24:25,913][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-22 17:24:25,913][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:24:25,920][metrics][INFO] - Evaluating forget_quality
+[2026-07-22 17:24:25,921][evaluator][INFO] - Result for metric forget_quality:	1.02724452984523e-14
+[2026-07-22 17:24:26,329][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.50it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.50it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.35it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:02,  4.38it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.42it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.43it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.40it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.50it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.54it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.54it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.51it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.49it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.73it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.53it/s]
+Evaluated 400 examples
+[2026-07-22 17:24:29,201][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.5495166015625
+[2026-07-22 17:24:29,555][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:02,  5.02it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:02,  5.20it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:02,  4.03it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:00<00:01,  5.42it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:01<00:00,  7.08it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:01<00:00,  8.12it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:01<00:00,  6.49it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:01<00:00,  6.39it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:01<00:00,  5.98it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:02<00:00,  6.39it/s]
+Evaluated 400 examples
+[2026-07-22 17:24:31,590][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.020769793233637722
+[2026-07-22 17:24:31,903][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.62it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.70it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.73it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.76it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.76it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.69it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.68it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.65it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.68it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.67it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.62it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.65it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.87it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.73it/s]
+Evaluated 400 examples
+[2026-07-22 17:24:35,019][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:03,  3.18it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:03,  3.39it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:03,  3.27it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:02,  3.35it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:02,  3.57it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:01<00:02,  3.35it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:02<00:01,  3.34it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  3.33it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:02<00:01,  3.34it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:03<00:00,  3.26it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:03<00:00,  3.00it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:03<00:00,  2.96it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:04<00:00,  3.11it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:04<00:00,  3.22it/s]
+Evaluated 400 examples
+[2026-07-22 17:24:39,359][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.49it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.50it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.49it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.58it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.63it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.58it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.55it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.51it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.51it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.53it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.50it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.53it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.77it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.59it/s]
+Evaluated 400 examples
+[2026-07-22 17:24:42,581][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:12,  1.06s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:11,  1.07s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:10,  1.08s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:09,  1.07s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:08,  1.07s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.08s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.09s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.09s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:09<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:10<00:03,  1.09s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:11<00:02,  1.10s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.10s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.06s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.08s/it]
+Evaluated 400 examples
+[2026-07-22 17:24:56,607][metrics][INFO] - Evaluating retain_Truth_Ratio
+[2026-07-22 17:24:57,013][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.21it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.58it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  7.60it/s]
+Evaluated 100 examples
+[2026-07-22 17:24:57,894][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.47it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.51it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.10it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.51it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.55it/s]
+Evaluated 100 examples
+[2026-07-22 17:24:59,467][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+[2026-07-22 17:24:59,768][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  5.38it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  5.01it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  5.68it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  6.43it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  6.00it/s]
+Evaluated 100 examples
+[2026-07-22 17:25:00,436][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-22 17:25:00,436][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:25:00,436][metrics][INFO] - Evaluating ra_Truth_Ratio
+[2026-07-22 17:25:00,793][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  7.96it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.46it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  8.20it/s]
+Evaluated 117 examples
+[2026-07-22 17:25:01,587][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.62it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.47it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.07it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.54it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.55it/s]
+Evaluated 117 examples
+[2026-07-22 17:25:03,154][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+[2026-07-22 17:25:03,459][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  3.93it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  3.93it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  3.75it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  3.72it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  3.77it/s]
+Evaluated 117 examples
+[2026-07-22 17:25:04,522][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-22 17:25:04,522][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:25:04,522][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-22 17:25:04,522][metrics][INFO] - Evaluating model_utility
+[2026-07-22 17:25:04,522][evaluator][INFO] - Result for metric model_utility:	0.5284420115998628
+[2026-07-22 17:25:05,183][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:25:05,190][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 14.16it/s]
 31%|███       | 4/13 [00:00<00:00, 12.88it/s]
 46%|████▌     | 6/13 [00:00<00:00, 13.50it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.88it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 14.29it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 14.01it/s]
100%|██████████| 13/13 [00:00<00:00, 14.48it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.74it/s]
 31%|███       | 4/13 [00:00<00:00, 13.58it/s]
 46%|████▌     | 6/13 [00:00<00:00, 13.98it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.97it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 14.02it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 14.09it/s]
100%|██████████| 13/13 [00:00<00:00, 14.45it/s]
+[2026-07-22 17:25:06,990][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:25:06,997][metrics][INFO] - Evaluating privleak
+[2026-07-22 17:25:06,997][evaluator][INFO] - Result for metric privleak:	-96.58287945632476
+[2026-07-22 17:25:07,344][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:00, 12.07it/s]
Calculating ES:  31%|███       | 4/13 [00:00<00:00, 10.92it/s]
Calculating ES:  46%|████▌     | 6/13 [00:00<00:00, 11.65it/s]
Calculating ES:  62%|██████▏   | 8/13 [00:00<00:00, 12.12it/s]
Calculating ES:  77%|███████▋  | 10/13 [00:00<00:00, 12.23it/s]
Calculating ES:  92%|█████████▏| 12/13 [00:01<00:00, 11.98it/s]
Calculating ES: 100%|██████████| 13/13 [00:01<00:00, 12.40it/s]
+Evaluated 400 examples
+[2026-07-22 17:25:08,393][evaluator][INFO] - Result for metric extraction_strength:	0.23153243063984383
diff --git a/logs/eval_NPO_lr1e-05_beta0.05_alpha1_epoch10.log b/logs/eval_NPO_lr1e-05_beta0.05_alpha1_epoch10.log
new file mode 100644
index 0000000000000000000000000000000000000000..df97d102e449e0c731a9921751f925abcafd242e
--- /dev/null
+++ b/logs/eval_NPO_lr1e-05_beta0.05_alpha1_epoch10.log
@@ -0,0 +1,124 @@
+[2026-07-21 07:51:54,514][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/NPO_lr1e-05_beta0.05_alpha1_epoch10
+[2026-07-21 07:51:54,516][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-21 07:51:54,516][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/NPO_lr1e-05_beta0.05_alpha1_epoch10/TOFU_EVAL.json
+[2026-07-21 07:51:54,516][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/NPO_lr1e-05_beta0.05_alpha1_epoch10/TOFU_SUMMARY.json
+[2026-07-21 07:51:55,163][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:08,  1.46it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:04,  2.46it/s]
Calculating loss:  23%|██▎       | 3/13 [00:01<00:03,  3.11it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.57it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:02,  3.94it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.18it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.37it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:02<00:01,  4.54it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.63it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.69it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.71it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.71it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.86it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.08it/s]
+Evaluated 400 examples
+[2026-07-21 07:51:58,724][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:12,  1.03s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:11,  1.03s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:10,  1.05s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:09,  1.05s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:08,  1.04s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.03s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.02s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.01s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:09<00:04,  1.01s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:10<00:03,  1.01s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:11<00:02,  1.02s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:12<00:01,  1.02s/it]
Calculating loss: 100%|██████████| 13/13 [00:13<00:00,  1.03it/s]
Calculating loss: 100%|██████████| 13/13 [00:13<00:00,  1.01s/it]
+Evaluated 400 examples
+[2026-07-21 07:52:11,870][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-21 07:52:11,871][evaluator][INFO] - Result for metric forget_truth_ratio:	0.6877360924998683
+[2026-07-21 07:52:11,879][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-21 07:52:11,881][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-21 07:52:11,897][metrics][INFO] - Evaluating forget_quality
+[2026-07-21 07:52:11,899][evaluator][INFO] - Result for metric forget_quality:	0.0013468720321565051
+[2026-07-21 07:52:12,292][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.79it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.80it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.67it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.70it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.73it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.77it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.77it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.85it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.88it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.86it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.84it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.82it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  5.08it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.86it/s]
+Evaluated 400 examples
+[2026-07-21 07:52:14,965][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.0645626974105835
+[2026-07-21 07:52:15,247][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:03,  3.06it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:03,  3.01it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:03,  3.27it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:02,  3.44it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:02,  3.74it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:03<00:06,  1.16it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:03<00:03,  1.51it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:03<00:02,  1.87it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:04<00:01,  2.26it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:04<00:01,  2.64it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:04<00:00,  2.61it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:05<00:00,  2.92it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:05<00:00,  3.02it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:05<00:00,  2.42it/s]
+Evaluated 400 examples
+[2026-07-21 07:52:20,627][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.13478649070874613
+[2026-07-21 07:52:20,930][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.94it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  5.01it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:01,  5.02it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  5.04it/s]
Calculating loss:  38%|███▊      | 5/13 [00:00<00:01,  5.04it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.97it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.96it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.93it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.96it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.95it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.90it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.94it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  5.18it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  5.02it/s]
+Using the latest cached version of the dataset since locuslab/TOFU couldn't be found on the Hugging Face Hub
+Evaluated 400 examples
+[2026-07-21 07:52:23,559][datasets.load][WARNING] - Using the latest cached version of the dataset since locuslab/TOFU couldn't be found on the Hugging Face Hub
+Found the latest cached dataset configuration 'retain_perturbed' at /workspace/hf_cache_ou/locuslab___tofu/retain_perturbed/0.0.0/324592d84ae4f482ac7249b9285c2ecdb53e3a68 (last modified on Mon Jul 20 13:22:39 2026).
+[2026-07-21 07:52:23,561][datasets.packaged_modules.cache.cache][WARNING] - Found the latest cached dataset configuration 'retain_perturbed' at /workspace/hf_cache_ou/locuslab___tofu/retain_perturbed/0.0.0/324592d84ae4f482ac7249b9285c2ecdb53e3a68 (last modified on Mon Jul 20 13:22:39 2026).
+[2026-07-21 07:52:23,569][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:03,  3.71it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:02,  4.23it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:02,  4.09it/s]
Calculating text similarity:  31%|███       | 4/13 [00:00<00:02,  4.21it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:02,  3.68it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:01<00:01,  3.77it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:01<00:01,  3.73it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  3.87it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:02<00:01,  3.98it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:02<00:00,  3.90it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:02<00:00,  3.71it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:03<00:00,  3.50it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.95it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.87it/s]
+Using the latest cached version of the dataset since locuslab/TOFU couldn't be found on the Hugging Face Hub
+Evaluated 400 examples
+[2026-07-21 07:52:26,971][datasets.load][WARNING] - Using the latest cached version of the dataset since locuslab/TOFU couldn't be found on the Hugging Face Hub
+Found the latest cached dataset configuration 'retain_perturbed' at /workspace/hf_cache_ou/locuslab___tofu/retain_perturbed/0.0.0/324592d84ae4f482ac7249b9285c2ecdb53e3a68 (last modified on Mon Jul 20 13:22:39 2026).
+[2026-07-21 07:52:26,974][datasets.packaged_modules.cache.cache][WARNING] - Found the latest cached dataset configuration 'retain_perturbed' at /workspace/hf_cache_ou/locuslab___tofu/retain_perturbed/0.0.0/324592d84ae4f482ac7249b9285c2ecdb53e3a68 (last modified on Mon Jul 20 13:22:39 2026).
+[2026-07-21 07:52:26,979][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.92it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.93it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.91it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.95it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.98it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.90it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.87it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.82it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.82it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.82it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.80it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.82it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  5.08it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.92it/s]
+Using the latest cached version of the dataset since locuslab/TOFU couldn't be found on the Hugging Face Hub
+Evaluated 400 examples
+[2026-07-21 07:52:29,666][datasets.load][WARNING] - Using the latest cached version of the dataset since locuslab/TOFU couldn't be found on the Hugging Face Hub
+Found the latest cached dataset configuration 'retain_perturbed' at /workspace/hf_cache_ou/locuslab___tofu/retain_perturbed/0.0.0/324592d84ae4f482ac7249b9285c2ecdb53e3a68 (last modified on Mon Jul 20 13:22:39 2026).
+[2026-07-21 07:52:29,668][datasets.packaged_modules.cache.cache][WARNING] - Found the latest cached dataset configuration 'retain_perturbed' at /workspace/hf_cache_ou/locuslab___tofu/retain_perturbed/0.0.0/324592d84ae4f482ac7249b9285c2ecdb53e3a68 (last modified on Mon Jul 20 13:22:39 2026).
+[2026-07-21 07:52:29,673][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:11,  1.02it/s]
Calculating loss:  15%|█▌        | 2/13 [00:01<00:10,  1.02it/s]
Calculating loss:  23%|██▎       | 3/13 [00:02<00:09,  1.01it/s]
Calculating loss:  31%|███       | 4/13 [00:03<00:08,  1.02it/s]
Calculating loss:  38%|███▊      | 5/13 [00:04<00:07,  1.03it/s]
Calculating loss:  46%|████▌     | 6/13 [00:05<00:06,  1.02it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:06<00:05,  1.01it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:07<00:05,  1.00s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:08<00:04,  1.00s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:09<00:03,  1.00s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:10<00:02,  1.01s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:11<00:01,  1.01s/it]
Calculating loss: 100%|██████████| 13/13 [00:12<00:00,  1.04it/s]
Calculating loss: 100%|██████████| 13/13 [00:12<00:00,  1.01it/s]
+Evaluated 400 examples
+[2026-07-21 07:52:42,488][metrics][INFO] - Evaluating retain_Truth_Ratio
+Using the latest cached version of the dataset since locuslab/TOFU couldn't be found on the Hugging Face Hub
+[2026-07-21 07:52:42,526][datasets.load][WARNING] - Using the latest cached version of the dataset since locuslab/TOFU couldn't be found on the Hugging Face Hub
+Found the latest cached dataset configuration 'real_authors_perturbed' at /workspace/hf_cache_ou/locuslab___tofu/real_authors_perturbed/0.0.0/324592d84ae4f482ac7249b9285c2ecdb53e3a68 (last modified on Mon Jul 20 13:23:07 2026).
+[2026-07-21 07:52:42,527][datasets.packaged_modules.cache.cache][WARNING] - Found the latest cached dataset configuration 'real_authors_perturbed' at /workspace/hf_cache_ou/locuslab___tofu/real_authors_perturbed/0.0.0/324592d84ae4f482ac7249b9285c2ecdb53e3a68 (last modified on Mon Jul 20 13:23:07 2026).
+[2026-07-21 07:52:42,533][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.43it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.85it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  7.99it/s]
+Using the latest cached version of the dataset since locuslab/TOFU couldn't be found on the Hugging Face Hub
+Evaluated 100 examples
+[2026-07-21 07:52:43,070][datasets.load][WARNING] - Using the latest cached version of the dataset since locuslab/TOFU couldn't be found on the Hugging Face Hub
+Found the latest cached dataset configuration 'real_authors_perturbed' at /workspace/hf_cache_ou/locuslab___tofu/real_authors_perturbed/0.0.0/324592d84ae4f482ac7249b9285c2ecdb53e3a68 (last modified on Mon Jul 20 13:23:07 2026).
+[2026-07-21 07:52:43,072][datasets.packaged_modules.cache.cache][WARNING] - Found the latest cached dataset configuration 'real_authors_perturbed' at /workspace/hf_cache_ou/locuslab___tofu/real_authors_perturbed/0.0.0/324592d84ae4f482ac7249b9285c2ecdb53e3a68 (last modified on Mon Jul 20 13:23:07 2026).
+[2026-07-21 07:52:43,078][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.39it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.63it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.22it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.64it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.67it/s]
+Evaluated 100 examples
+[2026-07-21 07:52:44,577][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+Using the latest cached version of the dataset since locuslab/TOFU couldn't be found on the Hugging Face Hub
+[2026-07-21 07:52:44,612][datasets.load][WARNING] - Using the latest cached version of the dataset since locuslab/TOFU couldn't be found on the Hugging Face Hub
+Found the latest cached dataset configuration 'real_authors_perturbed' at /workspace/hf_cache_ou/locuslab___tofu/real_authors_perturbed/0.0.0/324592d84ae4f482ac7249b9285c2ecdb53e3a68 (last modified on Mon Jul 20 13:23:07 2026).
+[2026-07-21 07:52:44,613][datasets.packaged_modules.cache.cache][WARNING] - Found the latest cached dataset configuration 'real_authors_perturbed' at /workspace/hf_cache_ou/locuslab___tofu/real_authors_perturbed/0.0.0/324592d84ae4f482ac7249b9285c2ecdb53e3a68 (last modified on Mon Jul 20 13:23:07 2026).
+[2026-07-21 07:52:44,617][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  5.05it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  4.22it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  4.76it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  5.88it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  5.34it/s]
+Evaluated 100 examples
+[2026-07-21 07:52:45,367][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-21 07:52:45,368][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-21 07:52:45,368][metrics][INFO] - Evaluating ra_Truth_Ratio
+Using the latest cached version of the dataset since locuslab/TOFU couldn't be found on the Hugging Face Hub
+[2026-07-21 07:52:45,407][datasets.load][WARNING] - Using the latest cached version of the dataset since locuslab/TOFU couldn't be found on the Hugging Face Hub
+Found the latest cached dataset configuration 'world_facts_perturbed' at /workspace/hf_cache_ou/locuslab___tofu/world_facts_perturbed/0.0.0/324592d84ae4f482ac7249b9285c2ecdb53e3a68 (last modified on Mon Jul 20 13:23:12 2026).
+[2026-07-21 07:52:45,410][datasets.packaged_modules.cache.cache][WARNING] - Found the latest cached dataset configuration 'world_facts_perturbed' at /workspace/hf_cache_ou/locuslab___tofu/world_facts_perturbed/0.0.0/324592d84ae4f482ac7249b9285c2ecdb53e3a68 (last modified on Mon Jul 20 13:23:12 2026).
+[2026-07-21 07:52:45,414][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.38it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.90it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  8.67it/s]
+Using the latest cached version of the dataset since locuslab/TOFU couldn't be found on the Hugging Face Hub
+Evaluated 117 examples
+[2026-07-21 07:52:45,911][datasets.load][WARNING] - Using the latest cached version of the dataset since locuslab/TOFU couldn't be found on the Hugging Face Hub
+Found the latest cached dataset configuration 'world_facts_perturbed' at /workspace/hf_cache_ou/locuslab___tofu/world_facts_perturbed/0.0.0/324592d84ae4f482ac7249b9285c2ecdb53e3a68 (last modified on Mon Jul 20 13:23:12 2026).
+[2026-07-21 07:52:45,912][datasets.packaged_modules.cache.cache][WARNING] - Found the latest cached dataset configuration 'world_facts_perturbed' at /workspace/hf_cache_ou/locuslab___tofu/world_facts_perturbed/0.0.0/324592d84ae4f482ac7249b9285c2ecdb53e3a68 (last modified on Mon Jul 20 13:23:12 2026).
+[2026-07-21 07:52:45,915][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.54it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.62it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.22it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.71it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.72it/s]
+Evaluated 117 examples
+[2026-07-21 07:52:47,388][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+Using the latest cached version of the dataset since locuslab/TOFU couldn't be found on the Hugging Face Hub
+[2026-07-21 07:52:47,423][datasets.load][WARNING] - Using the latest cached version of the dataset since locuslab/TOFU couldn't be found on the Hugging Face Hub
+Found the latest cached dataset configuration 'world_facts_perturbed' at /workspace/hf_cache_ou/locuslab___tofu/world_facts_perturbed/0.0.0/324592d84ae4f482ac7249b9285c2ecdb53e3a68 (last modified on Mon Jul 20 13:23:12 2026).
+[2026-07-21 07:52:47,425][datasets.packaged_modules.cache.cache][WARNING] - Found the latest cached dataset configuration 'world_facts_perturbed' at /workspace/hf_cache_ou/locuslab___tofu/world_facts_perturbed/0.0.0/324592d84ae4f482ac7249b9285c2ecdb53e3a68 (last modified on Mon Jul 20 13:23:12 2026).
+[2026-07-21 07:52:47,429][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  5.09it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  3.61it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  3.91it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  4.13it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  4.08it/s]
+Evaluated 117 examples
+[2026-07-21 07:52:48,411][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-21 07:52:48,411][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-21 07:52:48,411][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-21 07:52:48,411][metrics][INFO] - Evaluating model_utility
+[2026-07-21 07:52:48,412][evaluator][INFO] - Result for metric model_utility:	0.2596283949532122
+Using the latest cached version of the dataset since locuslab/TOFU couldn't be found on the Hugging Face Hub
+[2026-07-21 07:52:48,485][datasets.load][WARNING] - Using the latest cached version of the dataset since locuslab/TOFU couldn't be found on the Hugging Face Hub
+Found the latest cached dataset configuration 'forget10_perturbed' at /workspace/hf_cache_ou/locuslab___tofu/forget10_perturbed/0.0.0/324592d84ae4f482ac7249b9285c2ecdb53e3a68 (last modified on Mon Jul 20 13:17:15 2026).
+[2026-07-21 07:52:48,487][datasets.packaged_modules.cache.cache][WARNING] - Found the latest cached dataset configuration 'forget10_perturbed' at /workspace/hf_cache_ou/locuslab___tofu/forget10_perturbed/0.0.0/324592d84ae4f482ac7249b9285c2ecdb53e3a68 (last modified on Mon Jul 20 13:17:15 2026).
+Using the latest cached version of the dataset since locuslab/TOFU couldn't be found on the Hugging Face Hub
+[2026-07-21 07:52:48,528][datasets.load][WARNING] - Using the latest cached version of the dataset since locuslab/TOFU couldn't be found on the Hugging Face Hub
+Found the latest cached dataset configuration 'holdout10' at /workspace/hf_cache_ou/locuslab___tofu/holdout10/0.0.0/324592d84ae4f482ac7249b9285c2ecdb53e3a68 (last modified on Mon Jul 20 13:23:17 2026).
+[2026-07-21 07:52:48,531][datasets.packaged_modules.cache.cache][WARNING] - Found the latest cached dataset configuration 'holdout10' at /workspace/hf_cache_ou/locuslab___tofu/holdout10/0.0.0/324592d84ae4f482ac7249b9285c2ecdb53e3a68 (last modified on Mon Jul 20 13:23:17 2026).
+[2026-07-21 07:52:48,534][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-21 07:52:48,550][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.28it/s]
 31%|███       | 4/13 [00:00<00:00, 11.41it/s]
 46%|████▌     | 6/13 [00:00<00:00, 11.70it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 11.86it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 12.10it/s]
 92%|█████████▏| 12/13 [00:01<00:00, 11.86it/s]
100%|██████████| 13/13 [00:01<00:00, 12.39it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 11.50it/s]
 31%|███       | 4/13 [00:00<00:00, 11.38it/s]
 46%|████▌     | 6/13 [00:00<00:00, 11.70it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 11.76it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 11.77it/s]
 92%|█████████▏| 12/13 [00:01<00:00, 11.89it/s]
100%|██████████| 13/13 [00:01<00:00, 12.18it/s]
+[2026-07-21 07:52:50,674][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-21 07:52:50,689][metrics][INFO] - Evaluating privleak
+[2026-07-21 07:52:50,690][evaluator][INFO] - Result for metric privleak:	5.302079325334236
+Using the latest cached version of the dataset since locuslab/TOFU couldn't be found on the Hugging Face Hub
+[2026-07-21 07:52:50,764][datasets.load][WARNING] - Using the latest cached version of the dataset since locuslab/TOFU couldn't be found on the Hugging Face Hub
+Found the latest cached dataset configuration 'forget10_perturbed' at /workspace/hf_cache_ou/locuslab___tofu/forget10_perturbed/0.0.0/324592d84ae4f482ac7249b9285c2ecdb53e3a68 (last modified on Mon Jul 20 13:17:15 2026).
+[2026-07-21 07:52:50,766][datasets.packaged_modules.cache.cache][WARNING] - Found the latest cached dataset configuration 'forget10_perturbed' at /workspace/hf_cache_ou/locuslab___tofu/forget10_perturbed/0.0.0/324592d84ae4f482ac7249b9285c2ecdb53e3a68 (last modified on Mon Jul 20 13:17:15 2026).
+[2026-07-21 07:52:50,772][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:01, 10.05it/s]
Calculating ES:  31%|███       | 4/13 [00:00<00:00,  9.17it/s]
Calculating ES:  38%|███▊      | 5/13 [00:00<00:00,  9.36it/s]
Calculating ES:  54%|█████▍    | 7/13 [00:00<00:00,  9.62it/s]
Calculating ES:  69%|██████▉   | 9/13 [00:00<00:00, 10.03it/s]
Calculating ES:  85%|████████▍ | 11/13 [00:01<00:00,  9.88it/s]
Calculating ES:  92%|█████████▏| 12/13 [00:01<00:00,  9.74it/s]
Calculating ES: 100%|██████████| 13/13 [00:01<00:00, 10.10it/s]
+Evaluated 400 examples
+[2026-07-21 07:52:52,060][evaluator][INFO] - Result for metric extraction_strength:	0.06207448107554553
diff --git a/logs/eval_NPO_lr1e-05_beta0.05_alpha1_epoch5.log b/logs/eval_NPO_lr1e-05_beta0.05_alpha1_epoch5.log
new file mode 100644
index 0000000000000000000000000000000000000000..9f99e03e32ea52bb2ddf497df25ff7643ad17ba6
--- /dev/null
+++ b/logs/eval_NPO_lr1e-05_beta0.05_alpha1_epoch5.log
@@ -0,0 +1,78 @@
+/opt/conda/envs/copygen/lib/python3.12/site-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.6.3) or chardet (6.0.0.post1)/charset_normalizer (3.4.4) doesn't match a supported version!
+  warnings.warn(
+[2026-07-22 17:17:36,998][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/NPO_lr1e-05_beta0.05_alpha1_epoch5
+[2026-07-22 17:17:36,999][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-22 17:17:36,999][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/NPO_lr1e-05_beta0.05_alpha1_epoch5/TOFU_EVAL.json
+[2026-07-22 17:17:36,999][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/NPO_lr1e-05_beta0.05_alpha1_epoch5/TOFU_SUMMARY.json
+[2026-07-22 17:17:39,825][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:05,  2.33it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:03,  3.22it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  3.60it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.82it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.01it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.12it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.20it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:02<00:01,  4.31it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.36it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.36it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.32it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.29it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.48it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.11it/s]
+Evaluated 400 examples
+[2026-07-22 17:17:43,343][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:13,  1.12s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:12,  1.12s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:11,  1.14s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:10,  1.15s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:09,  1.14s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.13s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.13s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:09<00:05,  1.11s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:10<00:04,  1.11s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:11<00:03,  1.11s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:12<00:02,  1.12s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.12s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.07s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.11s/it]
+Evaluated 400 examples
+[2026-07-22 17:17:57,769][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-22 17:17:57,769][evaluator][INFO] - Result for metric forget_truth_ratio:	0.7040837745521601
+[2026-07-22 17:17:57,774][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-22 17:17:57,774][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:17:57,782][metrics][INFO] - Evaluating forget_quality
+[2026-07-22 17:17:57,783][evaluator][INFO] - Result for metric forget_quality:	3.596289960343452e-05
+[2026-07-22 17:17:58,166][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.47it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.45it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.30it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:02,  4.29it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.28it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.33it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.33it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.44it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.50it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.51it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.49it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.48it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.73it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.48it/s]
+Evaluated 400 examples
+[2026-07-22 17:18:01,066][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.07519393920898437
+[2026-07-22 17:18:01,359][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:02,  5.71it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:01,  7.01it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:01,  7.07it/s]
Calculating text similarity:  31%|███       | 4/13 [00:00<00:01,  7.29it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:00<00:01,  7.68it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:00<00:00,  7.75it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:00<00:00,  7.83it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:01<00:00,  8.00it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:01<00:00,  7.93it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:01<00:00,  7.90it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:01<00:00,  7.68it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:01<00:00,  7.86it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:01<00:00,  8.24it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:01<00:00,  7.75it/s]
+Evaluated 400 examples
+[2026-07-22 17:18:03,037][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.12957986687563722
+[2026-07-22 17:18:03,284][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.66it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.64it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.67it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.70it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.70it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.58it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.59it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.58it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.62it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.61it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.56it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.60it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.83it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.67it/s]
+Evaluated 400 examples
+[2026-07-22 17:18:06,322][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:01,  7.15it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:01,  7.36it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:01,  8.01it/s]
Calculating text similarity:  31%|███       | 4/13 [00:00<00:01,  7.69it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:00<00:01,  7.65it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:00<00:00,  7.18it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:00<00:00,  7.02it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:01<00:00,  7.31it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:01<00:00,  7.36it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:01<00:00,  7.59it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:01<00:00,  7.14it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:01<00:00,  7.49it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:01<00:00,  7.65it/s]
+Evaluated 400 examples
+[2026-07-22 17:18:08,633][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.62it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.63it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.60it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.66it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.69it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.61it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.57it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.52it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.52it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.53it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.50it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.53it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.77it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.62it/s]
+Evaluated 400 examples
+[2026-07-22 17:18:11,724][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:12,  1.07s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:11,  1.07s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:10,  1.08s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:09,  1.08s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:08,  1.07s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.08s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.09s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.10s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:09<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:10<00:03,  1.10s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:12<00:02,  1.11s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.11s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.05s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.08s/it]
+Evaluated 400 examples
+[2026-07-22 17:18:25,763][metrics][INFO] - Evaluating retain_Truth_Ratio
+[2026-07-22 17:18:26,158][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.24it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.61it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  7.61it/s]
+Evaluated 100 examples
+[2026-07-22 17:18:26,965][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.56it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.51it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.08it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.44it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.50it/s]
+Evaluated 100 examples
+[2026-07-22 17:18:28,568][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+[2026-07-22 17:18:28,825][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  9.08it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  6.76it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  9.66it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  9.11it/s]
+Evaluated 100 examples
+[2026-07-22 17:18:29,265][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-22 17:18:29,265][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:18:29,265][metrics][INFO] - Evaluating ra_Truth_Ratio
+[2026-07-22 17:18:29,543][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  7.95it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.52it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  8.26it/s]
+Evaluated 117 examples
+[2026-07-22 17:18:30,282][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.45it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.47it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.08it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.53it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.55it/s]
+Evaluated 117 examples
+[2026-07-22 17:18:31,851][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+[2026-07-22 17:18:32,113][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  6.13it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  7.23it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  6.76it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  7.24it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  7.06it/s]
+Evaluated 117 examples
+[2026-07-22 17:18:32,680][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-22 17:18:32,680][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:18:32,680][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-22 17:18:32,681][metrics][INFO] - Evaluating model_utility
+[2026-07-22 17:18:32,681][evaluator][INFO] - Result for metric model_utility:	0.1728273832476991
+[2026-07-22 17:18:33,299][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:18:33,306][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 14.06it/s]
 31%|███       | 4/13 [00:00<00:00, 12.71it/s]
 46%|████▌     | 6/13 [00:00<00:00, 13.31it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.65it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 14.00it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.71it/s]
100%|██████████| 13/13 [00:00<00:00, 14.21it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.39it/s]
 31%|███       | 4/13 [00:00<00:00, 13.25it/s]
 46%|████▌     | 6/13 [00:00<00:00, 13.66it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.68it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.73it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.82it/s]
100%|██████████| 13/13 [00:00<00:00, 14.16it/s]
+[2026-07-22 17:18:35,141][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:18:35,147][metrics][INFO] - Evaluating privleak
+[2026-07-22 17:18:35,147][evaluator][INFO] - Result for metric privleak:	-30.390152047070888
+[2026-07-22 17:18:35,437][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:00, 11.66it/s]
Calculating ES:  31%|███       | 4/13 [00:00<00:00, 10.64it/s]
Calculating ES:  46%|████▌     | 6/13 [00:00<00:00, 11.28it/s]
Calculating ES:  62%|██████▏   | 8/13 [00:00<00:00, 11.68it/s]
Calculating ES:  77%|███████▋  | 10/13 [00:00<00:00, 11.70it/s]
Calculating ES:  92%|█████████▏| 12/13 [00:01<00:00, 11.33it/s]
Calculating ES: 100%|██████████| 13/13 [00:01<00:00, 11.83it/s]
+Evaluated 400 examples
+[2026-07-22 17:18:36,537][evaluator][INFO] - Result for metric extraction_strength:	0.057134989491472735
diff --git a/logs/eval_NPO_lr1e-05_beta0.05_alpha2_epoch10.log b/logs/eval_NPO_lr1e-05_beta0.05_alpha2_epoch10.log
new file mode 100644
index 0000000000000000000000000000000000000000..87b3860c34036a06b66964e59255cb2c52901da5
--- /dev/null
+++ b/logs/eval_NPO_lr1e-05_beta0.05_alpha2_epoch10.log
@@ -0,0 +1,78 @@
+/opt/conda/envs/copygen/lib/python3.12/site-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.6.3) or chardet (6.0.0.post1)/charset_normalizer (3.4.4) doesn't match a supported version!
+  warnings.warn(
+[2026-07-22 17:25:36,309][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/NPO_lr1e-05_beta0.05_alpha2_epoch10
+[2026-07-22 17:25:36,310][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-22 17:25:36,310][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/NPO_lr1e-05_beta0.05_alpha2_epoch10/TOFU_EVAL.json
+[2026-07-22 17:25:36,310][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/NPO_lr1e-05_beta0.05_alpha2_epoch10/TOFU_SUMMARY.json
+[2026-07-22 17:25:37,119][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:05,  2.36it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:03,  3.23it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  3.59it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.76it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:02,  3.95it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.05it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.16it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:02<00:01,  4.29it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.35it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.40it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.40it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.39it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.60it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.13it/s]
+Evaluated 400 examples
+[2026-07-22 17:25:40,815][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:13,  1.11s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:12,  1.12s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:11,  1.15s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:10,  1.15s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:09,  1.14s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.14s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.13s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.11s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:10<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:11<00:03,  1.10s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:12<00:02,  1.11s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.12s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.07s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.11s/it]
+Evaluated 400 examples
+[2026-07-22 17:25:55,219][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-22 17:25:55,220][evaluator][INFO] - Result for metric forget_truth_ratio:	0.6744537705504727
+[2026-07-22 17:25:55,224][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-22 17:25:55,224][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:25:55,231][metrics][INFO] - Evaluating forget_quality
+[2026-07-22 17:25:55,231][evaluator][INFO] - Result for metric forget_quality:	0.0004368337391825136
+[2026-07-22 17:25:55,685][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.50it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.50it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.35it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:02,  4.38it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.42it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.47it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.46it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.55it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.59it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.58it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.55it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.53it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.77it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.56it/s]
+Evaluated 400 examples
+[2026-07-22 17:25:58,535][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.08624565601348877
+[2026-07-22 17:25:59,061][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:02,  4.33it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:02,  4.84it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:02,  4.94it/s]
Calculating text similarity:  31%|███       | 4/13 [00:00<00:01,  5.60it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:00<00:01,  5.79it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:01<00:01,  5.80it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:01<00:01,  5.57it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:01<00:00,  5.82it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:01<00:00,  5.96it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:01<00:00,  5.77it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:01<00:00,  5.48it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:02<00:00,  5.88it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:02<00:00,  5.94it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:02<00:00,  5.66it/s]
+Evaluated 400 examples
+[2026-07-22 17:26:01,360][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.1558186962244796
+[2026-07-22 17:26:01,699][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.61it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.69it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.69it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.72it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.64it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.56it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.55it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.53it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.59it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.61it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.57it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.62it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.85it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.67it/s]
+Evaluated 400 examples
+[2026-07-22 17:26:04,923][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:02,  5.70it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:01,  6.32it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:01,  6.21it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:04,  1.96it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:03,  2.60it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:02<00:04,  1.59it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:02<00:02,  2.06it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:03<00:01,  2.53it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:03<00:01,  3.12it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:03<00:00,  3.70it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:03<00:00,  4.26it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:03<00:00,  4.55it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  5.44it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.30it/s]
+Evaluated 400 examples
+[2026-07-22 17:26:09,171][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.61it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.62it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.58it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.63it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.65it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.58it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.52it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.42it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.41it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.42it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.41it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.45it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.71it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.56it/s]
+Evaluated 400 examples
+[2026-07-22 17:26:12,320][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:12,  1.06s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:11,  1.06s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:10,  1.08s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:09,  1.07s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:08,  1.06s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.08s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.10s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.10s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:09<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:10<00:03,  1.09s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:11<00:02,  1.10s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.10s/it]
Calculating loss: 100%|██████████| 13/13 [00:13<00:00,  1.05s/it]
Calculating loss: 100%|██████████| 13/13 [00:13<00:00,  1.08s/it]
+Evaluated 400 examples
+[2026-07-22 17:26:26,321][metrics][INFO] - Evaluating retain_Truth_Ratio
+[2026-07-22 17:26:26,774][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.03it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.42it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  7.40it/s]
+Evaluated 100 examples
+[2026-07-22 17:26:27,698][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  5.89it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.45it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.08it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.49it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.51it/s]
+Evaluated 100 examples
+[2026-07-22 17:26:29,291][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+[2026-07-22 17:26:29,600][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  6.97it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  7.26it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  7.98it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  8.60it/s]
+Evaluated 100 examples
+[2026-07-22 17:26:30,065][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-22 17:26:30,066][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:26:30,066][metrics][INFO] - Evaluating ra_Truth_Ratio
+[2026-07-22 17:26:30,396][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.19it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.65it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  8.45it/s]
+Evaluated 117 examples
+[2026-07-22 17:26:31,170][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.91it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.51it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.09it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.55it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.57it/s]
+Evaluated 117 examples
+[2026-07-22 17:26:32,724][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+[2026-07-22 17:26:33,098][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  6.77it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  7.49it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  5.88it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  6.38it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  6.44it/s]
+Evaluated 117 examples
+[2026-07-22 17:26:33,730][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-22 17:26:33,730][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:26:33,730][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-22 17:26:33,731][metrics][INFO] - Evaluating model_utility
+[2026-07-22 17:26:33,731][evaluator][INFO] - Result for metric model_utility:	0.32742401489794026
+[2026-07-22 17:26:34,379][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:26:34,389][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.68it/s]
 31%|███       | 4/13 [00:00<00:00, 12.45it/s]
 46%|████▌     | 6/13 [00:00<00:00, 13.15it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.34it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.52it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.08it/s]
100%|██████████| 13/13 [00:00<00:00, 13.72it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.39it/s]
 31%|███       | 4/13 [00:00<00:00, 13.20it/s]
 46%|████▌     | 6/13 [00:00<00:00, 13.27it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.21it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.28it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.35it/s]
100%|██████████| 13/13 [00:00<00:00, 13.74it/s]
+[2026-07-22 17:26:36,285][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:26:36,292][metrics][INFO] - Evaluating privleak
+[2026-07-22 17:26:36,293][evaluator][INFO] - Result for metric privleak:	0.9571581865591441
+[2026-07-22 17:26:36,636][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:00, 11.64it/s]
Calculating ES:  31%|███       | 4/13 [00:00<00:00, 10.39it/s]
Calculating ES:  46%|████▌     | 6/13 [00:00<00:00, 11.16it/s]
Calculating ES:  62%|██████▏   | 8/13 [00:00<00:00, 11.68it/s]
Calculating ES:  77%|███████▋  | 10/13 [00:00<00:00, 11.83it/s]
Calculating ES:  92%|█████████▏| 12/13 [00:01<00:00, 11.64it/s]
Calculating ES: 100%|██████████| 13/13 [00:01<00:00, 11.98it/s]
+Evaluated 400 examples
+[2026-07-22 17:26:37,722][evaluator][INFO] - Result for metric extraction_strength:	0.06894511800369839
diff --git a/logs/eval_NPO_lr1e-05_beta0.05_alpha2_epoch5.log b/logs/eval_NPO_lr1e-05_beta0.05_alpha2_epoch5.log
new file mode 100644
index 0000000000000000000000000000000000000000..abdde749fbaaab1b994533b6ba72a9e121b21143
--- /dev/null
+++ b/logs/eval_NPO_lr1e-05_beta0.05_alpha2_epoch5.log
@@ -0,0 +1,78 @@
+/opt/conda/envs/copygen/lib/python3.12/site-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.6.3) or chardet (6.0.0.post1)/charset_normalizer (3.4.4) doesn't match a supported version!
+  warnings.warn(
+[2026-07-22 17:22:46,720][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/NPO_lr1e-05_beta0.05_alpha2_epoch5
+[2026-07-22 17:22:46,721][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-22 17:22:46,721][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/NPO_lr1e-05_beta0.05_alpha2_epoch5/TOFU_EVAL.json
+[2026-07-22 17:22:46,721][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/NPO_lr1e-05_beta0.05_alpha2_epoch5/TOFU_SUMMARY.json
+[2026-07-22 17:22:47,457][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:05,  2.21it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:03,  3.13it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  3.54it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.80it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.01it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.12it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.23it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:02<00:01,  4.35it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.40it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.44it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.41it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.40it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.60it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.14it/s]
+Evaluated 400 examples
+[2026-07-22 17:22:51,117][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:13,  1.12s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:12,  1.11s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:11,  1.14s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:10,  1.14s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:09,  1.13s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.12s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.12s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.11s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:10<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:11<00:03,  1.10s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:12<00:02,  1.11s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.11s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.06s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.10s/it]
+Evaluated 400 examples
+[2026-07-22 17:23:05,434][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-22 17:23:05,435][evaluator][INFO] - Result for metric forget_truth_ratio:	0.6875118993975735
+[2026-07-22 17:23:05,439][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-22 17:23:05,439][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:23:05,446][metrics][INFO] - Evaluating forget_quality
+[2026-07-22 17:23:05,447][evaluator][INFO] - Result for metric forget_quality:	0.0001305755477065129
+[2026-07-22 17:23:05,865][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.35it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.43it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.29it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:02,  4.33it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.35it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.40it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.36it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.44it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.48it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.49it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.49it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.48it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.74it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.49it/s]
+Evaluated 400 examples
+[2026-07-22 17:23:08,758][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.09350917816162109
+[2026-07-22 17:23:09,116][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:01,  6.08it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:01,  6.47it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:01,  5.38it/s]
Calculating text similarity:  31%|███       | 4/13 [00:00<00:01,  6.31it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:00<00:01,  6.96it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:00<00:00,  7.61it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:01<00:00,  7.36it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:01<00:00,  7.72it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:01<00:00,  7.10it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:01<00:00,  7.35it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:01<00:00,  6.72it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:01<00:00,  6.85it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:01<00:00,  7.04it/s]
+Evaluated 400 examples
+[2026-07-22 17:23:10,962][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.12072863328457346
+[2026-07-22 17:23:11,263][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.69it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.74it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.71it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.74it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.74it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.68it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.66it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.63it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.67it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.66it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.60it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.64it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.86it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.72it/s]
+Evaluated 400 examples
+[2026-07-22 17:23:14,318][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:02,  5.28it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:01,  6.29it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:01,  6.85it/s]
Calculating text similarity:  31%|███       | 4/13 [00:00<00:01,  7.02it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:00<00:01,  6.36it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:00<00:01,  6.60it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:01<00:00,  6.87it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:01<00:00,  6.90it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:01<00:00,  6.85it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:01<00:00,  6.83it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:01<00:00,  7.09it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:01<00:00,  6.78it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:01<00:00,  6.92it/s]
+Evaluated 400 examples
+[2026-07-22 17:23:16,517][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.56it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.61it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.59it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.65it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.68it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.61it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.57it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.52it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.51it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.52it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.50it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.54it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.79it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.62it/s]
+Evaluated 400 examples
+[2026-07-22 17:23:19,670][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:12,  1.06s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:11,  1.06s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:10,  1.07s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:09,  1.06s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:08,  1.07s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.08s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.09s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.09s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:09<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:10<00:03,  1.09s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:11<00:02,  1.10s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.11s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.05s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.08s/it]
+Evaluated 400 examples
+[2026-07-22 17:23:33,676][metrics][INFO] - Evaluating retain_Truth_Ratio
+[2026-07-22 17:23:34,089][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.12it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.57it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  7.57it/s]
+Evaluated 100 examples
+[2026-07-22 17:23:34,904][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.47it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.52it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.10it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.50it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.54it/s]
+Evaluated 100 examples
+[2026-07-22 17:23:36,479][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+[2026-07-22 17:23:36,802][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  7.79it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  7.43it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  8.39it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  9.55it/s]
+Evaluated 100 examples
+[2026-07-22 17:23:37,222][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-22 17:23:37,222][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:23:37,222][metrics][INFO] - Evaluating ra_Truth_Ratio
+[2026-07-22 17:23:37,575][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.10it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.59it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  8.35it/s]
+Evaluated 117 examples
+[2026-07-22 17:23:38,346][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.64it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.49it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.09it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.56it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.57it/s]
+Evaluated 117 examples
+[2026-07-22 17:23:39,900][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+[2026-07-22 17:23:40,202][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  7.82it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  7.98it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  7.58it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  7.46it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  7.56it/s]
+Evaluated 117 examples
+[2026-07-22 17:23:40,731][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-22 17:23:40,731][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:23:40,731][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-22 17:23:40,732][metrics][INFO] - Evaluating model_utility
+[2026-07-22 17:23:40,732][evaluator][INFO] - Result for metric model_utility:	0.18557101233951934
+[2026-07-22 17:23:41,460][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:23:41,470][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.09it/s]
 31%|███       | 4/13 [00:00<00:00, 11.98it/s]
 46%|████▌     | 6/13 [00:00<00:00, 12.66it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.06it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.45it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.23it/s]
100%|██████████| 13/13 [00:00<00:00, 13.62it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.03it/s]
 31%|███       | 4/13 [00:00<00:00, 12.85it/s]
 46%|████▌     | 6/13 [00:00<00:00, 13.24it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.23it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.23it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.37it/s]
100%|██████████| 13/13 [00:00<00:00, 13.71it/s]
+[2026-07-22 17:23:43,376][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:23:43,382][metrics][INFO] - Evaluating privleak
+[2026-07-22 17:23:43,382][evaluator][INFO] - Result for metric privleak:	-35.686663556407154
+[2026-07-22 17:23:43,693][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:00, 11.50it/s]
Calculating ES:  31%|███       | 4/13 [00:00<00:00, 10.52it/s]
Calculating ES:  46%|████▌     | 6/13 [00:00<00:00, 11.19it/s]
Calculating ES:  62%|██████▏   | 8/13 [00:00<00:00, 11.63it/s]
Calculating ES:  77%|███████▋  | 10/13 [00:00<00:00, 11.80it/s]
Calculating ES:  92%|█████████▏| 12/13 [00:01<00:00, 11.59it/s]
Calculating ES: 100%|██████████| 13/13 [00:01<00:00, 11.96it/s]
+Evaluated 400 examples
+[2026-07-22 17:23:44,781][evaluator][INFO] - Result for metric extraction_strength:	0.06249982300166865
diff --git a/logs/eval_NPO_lr1e-05_beta0.05_alpha5_epoch10.log b/logs/eval_NPO_lr1e-05_beta0.05_alpha5_epoch10.log
new file mode 100644
index 0000000000000000000000000000000000000000..f539619717a0d172f2b4919b90b7454556a7ce67
--- /dev/null
+++ b/logs/eval_NPO_lr1e-05_beta0.05_alpha5_epoch10.log
@@ -0,0 +1,78 @@
+/opt/conda/envs/copygen/lib/python3.12/site-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.6.3) or chardet (6.0.0.post1)/charset_normalizer (3.4.4) doesn't match a supported version!
+  warnings.warn(
+[2026-07-22 17:19:03,914][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/NPO_lr1e-05_beta0.05_alpha5_epoch10
+[2026-07-22 17:19:03,915][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-22 17:19:03,915][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/NPO_lr1e-05_beta0.05_alpha5_epoch10/TOFU_EVAL.json
+[2026-07-22 17:19:03,915][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/NPO_lr1e-05_beta0.05_alpha5_epoch10/TOFU_SUMMARY.json
+[2026-07-22 17:19:04,622][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:05,  2.26it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:03,  3.15it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  3.52it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.74it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:02,  3.94it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.05it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.17it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:02<00:01,  4.32it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.36it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.41it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.40it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.39it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.59it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.12it/s]
+Evaluated 400 examples
+[2026-07-22 17:19:08,125][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:13,  1.12s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:12,  1.12s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:11,  1.15s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:10,  1.15s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:09,  1.14s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.14s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.13s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:09<00:05,  1.11s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:10<00:04,  1.11s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:11<00:03,  1.11s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:12<00:02,  1.12s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.12s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.07s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.11s/it]
+Evaluated 400 examples
+[2026-07-22 17:19:22,565][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-22 17:19:22,565][evaluator][INFO] - Result for metric forget_truth_ratio:	0.6292047787397415
+[2026-07-22 17:19:22,570][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-22 17:19:22,570][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:19:22,579][metrics][INFO] - Evaluating forget_quality
+[2026-07-22 17:19:22,580][evaluator][INFO] - Result for metric forget_quality:	0.0003254839625295495
+[2026-07-22 17:19:23,141][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.49it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.48it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.32it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:02,  4.36it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.40it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.45it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.45it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.53it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.57it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.55it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.53it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.50it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.74it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.54it/s]
+Evaluated 400 examples
+[2026-07-22 17:19:26,006][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.15761444091796875
+[2026-07-22 17:19:26,630][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:02,  4.46it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:02,  3.71it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:02,  4.03it/s]
Calculating text similarity:  31%|███       | 4/13 [00:00<00:01,  4.61it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:01,  4.67it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:01<00:01,  5.03it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:01<00:01,  4.83it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:01<00:01,  4.95it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:01<00:00,  5.01it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:02<00:00,  4.53it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:02<00:00,  4.43it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:02<00:00,  4.70it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:02<00:00,  4.44it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:02<00:00,  4.57it/s]
+Evaluated 400 examples
+[2026-07-22 17:19:29,477][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.15076141467060733
+[2026-07-22 17:19:30,172][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.63it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.66it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.57it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.59it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.60it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.53it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.56it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.55it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.60it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.60it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.55it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.58it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.81it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.63it/s]
+Evaluated 400 examples
+[2026-07-22 17:19:33,279][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:02,  4.45it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:02,  5.14it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:02,  4.13it/s]
Calculating text similarity:  31%|███       | 4/13 [00:00<00:02,  4.20it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:01,  4.55it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:01<00:01,  4.78it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:01<00:01,  4.93it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:01<00:00,  5.03it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:01<00:00,  4.75it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:02<00:00,  4.57it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:02<00:00,  4.15it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:02<00:00,  4.42it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:02<00:00,  4.99it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:02<00:00,  4.67it/s]
+Evaluated 400 examples
+[2026-07-22 17:19:36,352][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.47it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.53it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.52it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.59it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.63it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.53it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.51it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.46it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.47it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.49it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.46it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.49it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.74it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.57it/s]
+Evaluated 400 examples
+[2026-07-22 17:19:39,442][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:13,  1.09s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:11,  1.07s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:10,  1.08s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:09,  1.07s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:08,  1.06s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.08s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.09s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.10s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:09<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:10<00:03,  1.10s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:11<00:02,  1.10s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.10s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.05s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.08s/it]
+Evaluated 400 examples
+[2026-07-22 17:19:53,470][metrics][INFO] - Evaluating retain_Truth_Ratio
+[2026-07-22 17:19:53,839][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.15it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.55it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  7.49it/s]
+Evaluated 100 examples
+[2026-07-22 17:19:54,679][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.30it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.49it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.08it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.08it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.24it/s]
+Evaluated 100 examples
+[2026-07-22 17:19:56,465][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+[2026-07-22 17:19:56,703][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  7.17it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  7.65it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  7.81it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  9.27it/s]
+Evaluated 100 examples
+[2026-07-22 17:19:57,136][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-22 17:19:57,136][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:19:57,136][metrics][INFO] - Evaluating ra_Truth_Ratio
+[2026-07-22 17:19:57,402][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.18it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.64it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  8.44it/s]
+Evaluated 117 examples
+[2026-07-22 17:19:58,104][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.84it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.51it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.11it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.58it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.59it/s]
+Evaluated 117 examples
+[2026-07-22 17:19:59,646][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+[2026-07-22 17:19:59,892][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  9.06it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  8.86it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  7.13it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  7.67it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  7.80it/s]
+Evaluated 117 examples
+[2026-07-22 17:20:00,406][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-22 17:20:00,406][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:20:00,406][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-22 17:20:00,406][metrics][INFO] - Evaluating model_utility
+[2026-07-22 17:20:00,407][evaluator][INFO] - Result for metric model_utility:	0.4501295582236843
+[2026-07-22 17:20:00,963][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:20:00,970][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.63it/s]
 31%|███       | 4/13 [00:00<00:00, 12.38it/s]
 46%|████▌     | 6/13 [00:00<00:00, 12.95it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.25it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.60it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.31it/s]
100%|██████████| 13/13 [00:00<00:00, 13.78it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.15it/s]
 31%|███       | 4/13 [00:00<00:00, 13.01it/s]
 46%|████▌     | 6/13 [00:00<00:00, 13.36it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.18it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.33it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.44it/s]
100%|██████████| 13/13 [00:00<00:00, 13.79it/s]
+[2026-07-22 17:20:02,859][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:20:02,866][metrics][INFO] - Evaluating privleak
+[2026-07-22 17:20:02,866][evaluator][INFO] - Result for metric privleak:	-52.1324735191217
+[2026-07-22 17:20:03,126][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:01, 10.77it/s]
Calculating ES:  31%|███       | 4/13 [00:00<00:00,  9.83it/s]
Calculating ES:  46%|████▌     | 6/13 [00:00<00:00, 10.51it/s]
Calculating ES:  62%|██████▏   | 8/13 [00:00<00:00, 10.90it/s]
Calculating ES:  77%|███████▋  | 10/13 [00:00<00:00, 11.00it/s]
Calculating ES:  92%|█████████▏| 12/13 [00:01<00:00, 10.88it/s]
Calculating ES: 100%|██████████| 13/13 [00:01<00:00, 11.20it/s]
+Evaluated 400 examples
+[2026-07-22 17:20:04,287][evaluator][INFO] - Result for metric extraction_strength:	0.07698815361441316
diff --git a/logs/eval_NPO_lr1e-05_beta0.05_alpha5_epoch5.log b/logs/eval_NPO_lr1e-05_beta0.05_alpha5_epoch5.log
new file mode 100644
index 0000000000000000000000000000000000000000..912e29a48fe754b8caf53e2236db3ae975b0f445
--- /dev/null
+++ b/logs/eval_NPO_lr1e-05_beta0.05_alpha5_epoch5.log
@@ -0,0 +1,78 @@
+/opt/conda/envs/copygen/lib/python3.12/site-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.6.3) or chardet (6.0.0.post1)/charset_normalizer (3.4.4) doesn't match a supported version!
+  warnings.warn(
+[2026-07-22 17:27:04,381][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/NPO_lr1e-05_beta0.05_alpha5_epoch5
+[2026-07-22 17:27:04,382][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-22 17:27:04,382][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/NPO_lr1e-05_beta0.05_alpha5_epoch5/TOFU_EVAL.json
+[2026-07-22 17:27:04,382][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/NPO_lr1e-05_beta0.05_alpha5_epoch5/TOFU_SUMMARY.json
+[2026-07-22 17:27:05,164][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:07,  1.62it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:04,  2.58it/s]
Calculating loss:  23%|██▎       | 3/13 [00:01<00:03,  3.12it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.49it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:02,  3.78it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  3.95it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:02<00:01,  4.10it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:02<00:01,  4.25it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.32it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.36it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.34it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:03<00:00,  4.33it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.53it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  3.91it/s]
+Evaluated 400 examples
+[2026-07-22 17:27:08,955][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:13,  1.14s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:12,  1.13s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:11,  1.15s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:10,  1.15s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:09,  1.13s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.13s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.13s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:09<00:05,  1.11s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:10<00:04,  1.11s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:11<00:03,  1.10s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:12<00:02,  1.11s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.12s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.06s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.11s/it]
+Evaluated 400 examples
+[2026-07-22 17:27:23,347][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-22 17:27:23,348][evaluator][INFO] - Result for metric forget_truth_ratio:	0.6853083762251453
+[2026-07-22 17:27:23,352][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-22 17:27:23,352][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:27:23,359][metrics][INFO] - Evaluating forget_quality
+[2026-07-22 17:27:23,360][evaluator][INFO] - Result for metric forget_quality:	0.0001305755477065129
+[2026-07-22 17:27:23,786][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.47it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.44it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.31it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:02,  4.34it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.37it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.41it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.39it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.49it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.52it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.48it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.41it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.39it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.63it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.47it/s]
+Evaluated 400 examples
+[2026-07-22 17:27:26,696][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.1974298095703125
+[2026-07-22 17:27:27,006][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:02,  4.43it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:02,  4.66it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:01,  5.45it/s]
Calculating text similarity:  31%|███       | 4/13 [00:00<00:01,  5.89it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:00<00:01,  6.57it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:00<00:01,  6.84it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:01<00:00,  6.80it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:01<00:00,  7.07it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:01<00:00,  7.13it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:01<00:00,  6.51it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:01<00:00,  6.76it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:01<00:00,  7.18it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:01<00:00,  7.24it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:01<00:00,  6.59it/s]
+Evaluated 400 examples
+[2026-07-22 17:27:28,980][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.1313780968030782
+[2026-07-22 17:27:29,299][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.68it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.74it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.75it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.77it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.77it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.68it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.66it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.62it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.65it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.63it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.58it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.61it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.83it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.71it/s]
+Evaluated 400 examples
+[2026-07-22 17:27:32,370][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:01,  7.15it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:01,  7.87it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:01,  7.47it/s]
Calculating text similarity:  31%|███       | 4/13 [00:00<00:01,  7.66it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:00<00:01,  7.29it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:01<00:03,  2.09it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:01<00:02,  2.68it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  3.37it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:02<00:00,  4.09it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:02<00:00,  4.56it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:02<00:00,  4.64it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:02<00:00,  5.48it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:02<00:00,  4.67it/s]
+Evaluated 400 examples
+[2026-07-22 17:27:35,484][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.63it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.64it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.61it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.67it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.70it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.62it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.59it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.53it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.53it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.54it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.51it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.54it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.78it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.63it/s]
+Evaluated 400 examples
+[2026-07-22 17:27:38,614][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:12,  1.06s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:11,  1.06s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:10,  1.08s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:09,  1.08s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:08,  1.07s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.08s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.09s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.09s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:09<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:10<00:03,  1.09s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:11<00:02,  1.10s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.11s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.05s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.08s/it]
+Evaluated 400 examples
+[2026-07-22 17:27:52,633][metrics][INFO] - Evaluating retain_Truth_Ratio
+[2026-07-22 17:27:53,046][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.26it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.61it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  7.59it/s]
+Evaluated 100 examples
+[2026-07-22 17:27:53,872][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.59it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.50it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.09it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.45it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.50it/s]
+Evaluated 100 examples
+[2026-07-22 17:27:55,473][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+[2026-07-22 17:27:55,783][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  6.42it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  8.62it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  9.57it/s]
+Evaluated 100 examples
+[2026-07-22 17:27:56,201][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-22 17:27:56,202][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:27:56,202][metrics][INFO] - Evaluating ra_Truth_Ratio
+[2026-07-22 17:27:56,519][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.05it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.56it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  8.32it/s]
+Evaluated 117 examples
+[2026-07-22 17:27:57,356][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.45it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.48it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.08it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.53it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.55it/s]
+Evaluated 117 examples
+[2026-07-22 17:27:58,926][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+[2026-07-22 17:27:59,236][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  9.40it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  8.83it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  8.62it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  9.16it/s]
+Evaluated 117 examples
+[2026-07-22 17:27:59,674][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-22 17:27:59,674][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:27:59,674][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-22 17:27:59,674][metrics][INFO] - Evaluating model_utility
+[2026-07-22 17:27:59,674][evaluator][INFO] - Result for metric model_utility:	0.3557828895774852
+[2026-07-22 17:28:00,305][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:28:00,314][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.85it/s]
 31%|███       | 4/13 [00:00<00:00, 12.70it/s]
 46%|████▌     | 6/13 [00:00<00:00, 13.36it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.77it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 14.13it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.82it/s]
100%|██████████| 13/13 [00:00<00:00, 14.28it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.68it/s]
 31%|███       | 4/13 [00:00<00:00, 13.56it/s]
 46%|████▌     | 6/13 [00:00<00:00, 13.99it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 14.02it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 14.08it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 14.18it/s]
100%|██████████| 13/13 [00:00<00:00, 14.52it/s]
+[2026-07-22 17:28:02,123][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:28:02,129][metrics][INFO] - Evaluating privleak
+[2026-07-22 17:28:02,129][evaluator][INFO] - Result for metric privleak:	-68.06300741947918
+[2026-07-22 17:28:02,493][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:00, 11.74it/s]
Calculating ES:  31%|███       | 4/13 [00:00<00:00, 10.75it/s]
Calculating ES:  46%|████▌     | 6/13 [00:00<00:00, 11.32it/s]
Calculating ES:  62%|██████▏   | 8/13 [00:00<00:00, 11.66it/s]
Calculating ES:  77%|███████▋  | 10/13 [00:00<00:00, 11.77it/s]
Calculating ES:  92%|█████████▏| 12/13 [00:01<00:00, 11.36it/s]
Calculating ES: 100%|██████████| 13/13 [00:01<00:00, 11.87it/s]
+Evaluated 400 examples
+[2026-07-22 17:28:03,589][evaluator][INFO] - Result for metric extraction_strength:	0.07352854552860216
diff --git a/logs/eval_NPO_lr1e-05_beta0.1_alpha1_epoch10.log b/logs/eval_NPO_lr1e-05_beta0.1_alpha1_epoch10.log
new file mode 100644
index 0000000000000000000000000000000000000000..67c2116d2d2922478799bffb92e1f0d6925ce3fb
--- /dev/null
+++ b/logs/eval_NPO_lr1e-05_beta0.1_alpha1_epoch10.log
@@ -0,0 +1,78 @@
+/opt/conda/envs/copygen/lib/python3.12/site-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.6.3) or chardet (6.0.0.post1)/charset_normalizer (3.4.4) doesn't match a supported version!
+  warnings.warn(
+[2026-07-20 13:26:02,589][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/NPO_lr1e-05_beta0.1_alpha1_epoch10
+[2026-07-20 13:26:02,590][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-20 13:26:02,591][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/NPO_lr1e-05_beta0.1_alpha1_epoch10/TOFU_EVAL.json
+[2026-07-20 13:26:02,591][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/NPO_lr1e-05_beta0.1_alpha1_epoch10/TOFU_SUMMARY.json
+[2026-07-20 13:26:03,295][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:08,  1.42it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:04,  2.40it/s]
Calculating loss:  23%|██▎       | 3/13 [00:01<00:03,  2.85it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.34it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:02,  3.74it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.02it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:02<00:01,  4.22it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:02<00:01,  4.41it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.52it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.59it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.61it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:03<00:00,  4.62it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.77it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  3.94it/s]
+Evaluated 400 examples
+[2026-07-20 13:26:07,077][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:12,  1.03s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:11,  1.02s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:10,  1.04s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:09,  1.04s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:08,  1.03s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.03s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.02s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.01s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:09<00:04,  1.01s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:10<00:03,  1.01s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:11<00:02,  1.03s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:12<00:01,  1.03s/it]
Calculating loss: 100%|██████████| 13/13 [00:13<00:00,  1.03it/s]
Calculating loss: 100%|██████████| 13/13 [00:13<00:00,  1.01s/it]
+Evaluated 400 examples
+[2026-07-20 13:26:20,224][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-20 13:26:20,225][evaluator][INFO] - Result for metric forget_truth_ratio:	0.6409193995767771
+[2026-07-20 13:26:20,233][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-20 13:26:20,234][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-20 13:26:20,249][metrics][INFO] - Evaluating forget_quality
+[2026-07-20 13:26:20,251][evaluator][INFO] - Result for metric forget_quality:	0.00017795251634334
+[2026-07-20 13:26:20,616][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.80it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.70it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.61it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.67it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.71it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.76it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.77it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.85it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.88it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.49it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.58it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.64it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.94it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.76it/s]
+Evaluated 400 examples
+[2026-07-20 13:26:23,351][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.20775527954101564
+[2026-07-20 13:26:23,624][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:09,  1.33it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:01<00:06,  1.66it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:01<00:05,  1.77it/s]
Calculating text similarity:  31%|███       | 4/13 [00:02<00:05,  1.80it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:02<00:04,  1.98it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:03<00:03,  1.90it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:03<00:03,  1.98it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:04<00:02,  2.23it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:04<00:01,  2.28it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:05<00:01,  2.02it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:05<00:00,  2.08it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:06<00:00,  2.13it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:06<00:00,  2.10it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:06<00:00,  2.00it/s]
+Evaluated 400 examples
+[2026-07-20 13:26:30,124][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.18619367500997983
+[2026-07-20 13:26:30,535][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.94it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  5.02it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:01,  5.03it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  5.06it/s]
Calculating loss:  38%|███▊      | 5/13 [00:00<00:01,  5.06it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.99it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.98it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.93it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.97it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.96it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.92it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.96it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  5.19it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  5.03it/s]
+Evaluated 400 examples
+[2026-07-20 13:26:33,396][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:04,  2.44it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:04,  2.46it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:01<00:04,  2.31it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:04,  2.12it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:02<00:03,  2.30it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:02<00:03,  2.13it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:03<00:02,  2.01it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:03<00:02,  1.88it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:04<00:02,  1.86it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:04<00:01,  1.85it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:05<00:01,  1.83it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:06<00:00,  1.87it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:06<00:00,  1.57it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:06<00:00,  1.88it/s]
+Evaluated 400 examples
+[2026-07-20 13:26:40,673][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.88it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.90it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.83it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.90it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.93it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.87it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.84it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.80it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.79it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.81it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.79it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.82it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  5.08it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.90it/s]
+Evaluated 400 examples
+[2026-07-20 13:26:43,686][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:11,  1.02it/s]
Calculating loss:  15%|█▌        | 2/13 [00:01<00:10,  1.02it/s]
Calculating loss:  23%|██▎       | 3/13 [00:02<00:09,  1.00it/s]
Calculating loss:  31%|███       | 4/13 [00:03<00:08,  1.01it/s]
Calculating loss:  38%|███▊      | 5/13 [00:04<00:07,  1.02it/s]
Calculating loss:  46%|████▌     | 6/13 [00:05<00:06,  1.01it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:06<00:06,  1.00s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:07<00:05,  1.01s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:08<00:04,  1.01s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:09<00:03,  1.01s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:11<00:02,  1.02s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:12<00:01,  1.02s/it]
Calculating loss: 100%|██████████| 13/13 [00:12<00:00,  1.03it/s]
Calculating loss: 100%|██████████| 13/13 [00:12<00:00,  1.01it/s]
+Evaluated 400 examples
+[2026-07-20 13:26:56,590][metrics][INFO] - Evaluating retain_Truth_Ratio
+[2026-07-20 13:26:56,990][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.37it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.82it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  7.96it/s]
+Evaluated 100 examples
+[2026-07-20 13:26:57,760][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.23it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.61it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.18it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.61it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.64it/s]
+Evaluated 100 examples
+[2026-07-20 13:26:59,276][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+[2026-07-20 13:26:59,741][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:01,  2.71it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  2.62it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:01<00:00,  2.79it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  3.16it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  2.98it/s]
+Evaluated 100 examples
+[2026-07-20 13:27:01,084][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-20 13:27:01,085][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-20 13:27:01,085][metrics][INFO] - Evaluating ra_Truth_Ratio
+[2026-07-20 13:27:01,364][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.31it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.86it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  8.60it/s]
+Evaluated 117 examples
+[2026-07-20 13:27:02,086][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.41it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.61it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.21it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.69it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.70it/s]
+Evaluated 117 examples
+[2026-07-20 13:27:03,568][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+[2026-07-20 13:27:03,851][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:01,  2.23it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  2.64it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:01<00:00,  2.48it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  2.65it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  2.59it/s]
+Evaluated 117 examples
+[2026-07-20 13:27:05,400][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-20 13:27:05,400][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-20 13:27:05,400][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-20 13:27:05,401][metrics][INFO] - Evaluating model_utility
+[2026-07-20 13:27:05,402][evaluator][INFO] - Result for metric model_utility:	0.4312575503326488
+[2026-07-20 13:27:05,951][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-20 13:27:05,966][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.22it/s]
 31%|███       | 4/13 [00:00<00:00, 11.48it/s]
 46%|████▌     | 6/13 [00:00<00:00, 11.81it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 11.91it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 12.19it/s]
 92%|█████████▏| 12/13 [00:01<00:00, 11.89it/s]
100%|██████████| 13/13 [00:01<00:00, 12.45it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 11.61it/s]
 31%|███       | 4/13 [00:00<00:00, 11.52it/s]
 46%|████▌     | 6/13 [00:00<00:00, 11.88it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 11.17it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 10.97it/s]
 92%|█████████▏| 12/13 [00:01<00:00, 11.42it/s]
100%|██████████| 13/13 [00:01<00:00, 11.81it/s]
+[2026-07-20 13:27:08,116][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-20 13:27:08,132][metrics][INFO] - Evaluating privleak
+[2026-07-20 13:27:08,132][evaluator][INFO] - Result for metric privleak:	-52.45793767245727
+[2026-07-20 13:27:08,516][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:   8%|▊         | 1/13 [00:00<00:01,  9.49it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:01,  9.62it/s]
Calculating ES:  23%|██▎       | 3/13 [00:00<00:01,  8.95it/s]
Calculating ES:  31%|███       | 4/13 [00:00<00:00,  9.03it/s]
Calculating ES:  38%|███▊      | 5/13 [00:00<00:00,  9.32it/s]
Calculating ES:  54%|█████▍    | 7/13 [00:00<00:00,  9.71it/s]
Calculating ES:  69%|██████▉   | 9/13 [00:00<00:00, 10.13it/s]
Calculating ES:  77%|███████▋  | 10/13 [00:01<00:00, 10.03it/s]
Calculating ES:  85%|████████▍ | 11/13 [00:01<00:00,  9.96it/s]
Calculating ES:  92%|█████████▏| 12/13 [00:01<00:00,  9.82it/s]
Calculating ES: 100%|██████████| 13/13 [00:01<00:00, 10.11it/s]
+Evaluated 400 examples
+[2026-07-20 13:27:09,803][evaluator][INFO] - Result for metric extraction_strength:	0.09623633139113906
diff --git a/logs/eval_NPO_lr1e-05_beta0.1_alpha1_epoch5.log b/logs/eval_NPO_lr1e-05_beta0.1_alpha1_epoch5.log
new file mode 100644
index 0000000000000000000000000000000000000000..0f1f76593cb84adda4df51443fcae1074dd9890e
--- /dev/null
+++ b/logs/eval_NPO_lr1e-05_beta0.1_alpha1_epoch5.log
@@ -0,0 +1,78 @@
+/opt/conda/envs/copygen/lib/python3.12/site-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.6.3) or chardet (6.0.0.post1)/charset_normalizer (3.4.4) doesn't match a supported version!
+  warnings.warn(
+[2026-07-22 17:24:15,461][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/NPO_lr1e-05_beta0.1_alpha1_epoch5
+[2026-07-22 17:24:15,462][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-22 17:24:15,462][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/NPO_lr1e-05_beta0.1_alpha1_epoch5/TOFU_EVAL.json
+[2026-07-22 17:24:15,462][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/NPO_lr1e-05_beta0.1_alpha1_epoch5/TOFU_SUMMARY.json
+[2026-07-22 17:24:16,232][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:04,  2.46it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:03,  3.33it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  3.69it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.90it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.09it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.19it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.29it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.41it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.46it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.49it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.47it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.44it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.64it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.23it/s]
+Evaluated 400 examples
+[2026-07-22 17:24:19,727][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:13,  1.11s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:12,  1.11s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:11,  1.15s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:10,  1.15s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:09,  1.13s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.12s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.12s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.10s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:10<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:11<00:03,  1.10s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:12<00:02,  1.11s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.12s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.06s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.10s/it]
+Evaluated 400 examples
+[2026-07-22 17:24:34,060][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-22 17:24:34,060][evaluator][INFO] - Result for metric forget_truth_ratio:	0.6626269119139176
+[2026-07-22 17:24:34,064][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-22 17:24:34,064][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:24:34,071][metrics][INFO] - Evaluating forget_quality
+[2026-07-22 17:24:34,072][evaluator][INFO] - Result for metric forget_quality:	1.830133064154361e-05
+[2026-07-22 17:24:34,539][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.51it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.51it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.36it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:02,  4.39it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.42it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.44it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.43it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.53it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.57it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.56it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.53it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.51it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.75it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.55it/s]
+Evaluated 400 examples
+[2026-07-22 17:24:37,399][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.22589324951171874
+[2026-07-22 17:24:37,714][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:02,  4.15it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:02,  4.32it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:02,  3.61it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:02,  3.90it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:01,  4.28it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:01<00:01,  4.47it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:01<00:01,  4.17it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:01<00:01,  4.58it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:02<00:00,  4.66it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:02<00:00,  4.50it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:02<00:00,  4.67it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:02<00:00,  4.79it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:02<00:00,  4.64it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:02<00:00,  4.43it/s]
+Evaluated 400 examples
+[2026-07-22 17:24:40,649][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.15708466178816374
+[2026-07-22 17:24:40,960][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.68it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.75it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.76it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.79it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.79it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.72it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.71it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.67it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.70it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.69it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.64it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.67it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.90it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.76it/s]
+Evaluated 400 examples
+[2026-07-22 17:24:44,063][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:02,  4.88it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:02,  4.81it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:02,  4.75it/s]
Calculating text similarity:  31%|███       | 4/13 [00:00<00:01,  5.23it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:00<00:01,  5.47it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:01<00:01,  5.54it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:01<00:01,  5.65it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:01<00:00,  5.51it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:01<00:00,  5.58it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:01<00:00,  5.53it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:02<00:00,  5.11it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:02<00:00,  4.79it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:02<00:00,  5.00it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:02<00:00,  5.20it/s]
+Evaluated 400 examples
+[2026-07-22 17:24:46,877][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.63it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.63it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.50it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.54it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.57it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.52it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.51it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.48it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.49it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.52it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.50it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.53it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.78it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.59it/s]
+Evaluated 400 examples
+[2026-07-22 17:24:50,095][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:12,  1.06s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:11,  1.06s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:10,  1.07s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:09,  1.06s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:08,  1.06s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.08s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.09s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.09s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:09<00:04,  1.09s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:10<00:03,  1.09s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:11<00:02,  1.10s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.10s/it]
Calculating loss: 100%|██████████| 13/13 [00:13<00:00,  1.05s/it]
Calculating loss: 100%|██████████| 13/13 [00:13<00:00,  1.07s/it]
+Evaluated 400 examples
+[2026-07-22 17:25:04,060][metrics][INFO] - Evaluating retain_Truth_Ratio
+[2026-07-22 17:25:04,482][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.03it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.46it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  7.47it/s]
+Evaluated 100 examples
+[2026-07-22 17:25:05,608][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.36it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.51it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.10it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.52it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.55it/s]
+Evaluated 100 examples
+[2026-07-22 17:25:07,176][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+[2026-07-22 17:25:07,529][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  6.38it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  6.59it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  6.22it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  7.01it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  6.76it/s]
+Evaluated 100 examples
+[2026-07-22 17:25:08,122][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-22 17:25:08,122][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:25:08,122][metrics][INFO] - Evaluating ra_Truth_Ratio
+[2026-07-22 17:25:08,489][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.08it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.55it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  8.30it/s]
+Evaluated 117 examples
+[2026-07-22 17:25:09,282][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.55it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.47it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.08it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.55it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.56it/s]
+Evaluated 117 examples
+[2026-07-22 17:25:10,845][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+[2026-07-22 17:25:11,159][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  5.85it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  5.98it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  5.39it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  5.50it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  5.56it/s]
+Evaluated 117 examples
+[2026-07-22 17:25:11,879][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-22 17:25:11,879][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:25:11,879][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-22 17:25:11,879][metrics][INFO] - Evaluating model_utility
+[2026-07-22 17:25:11,880][evaluator][INFO] - Result for metric model_utility:	0.2935444678563043
+[2026-07-22 17:25:12,554][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:25:12,561][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.14it/s]
 31%|███       | 4/13 [00:00<00:00, 11.95it/s]
 46%|████▌     | 6/13 [00:00<00:00, 12.47it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 12.74it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.08it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 12.79it/s]
100%|██████████| 13/13 [00:00<00:00, 13.29it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 12.88it/s]
 31%|███       | 4/13 [00:00<00:00, 12.78it/s]
 46%|████▌     | 6/13 [00:00<00:00, 13.20it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.08it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 12.90it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.14it/s]
100%|██████████| 13/13 [00:00<00:00, 13.50it/s]
+[2026-07-22 17:25:14,505][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:25:14,512][metrics][INFO] - Evaluating privleak
+[2026-07-22 17:25:14,512][evaluator][INFO] - Result for metric privleak:	-71.68816179924468
+[2026-07-22 17:25:14,842][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:00, 11.55it/s]
Calculating ES:  31%|███       | 4/13 [00:00<00:00, 10.54it/s]
Calculating ES:  46%|████▌     | 6/13 [00:00<00:00, 11.17it/s]
Calculating ES:  62%|██████▏   | 8/13 [00:00<00:00, 11.57it/s]
Calculating ES:  77%|███████▋  | 10/13 [00:00<00:00, 11.71it/s]
Calculating ES:  92%|█████████▏| 12/13 [00:01<00:00, 11.48it/s]
Calculating ES: 100%|██████████| 13/13 [00:01<00:00, 11.88it/s]
+Evaluated 400 examples
+[2026-07-22 17:25:15,937][evaluator][INFO] - Result for metric extraction_strength:	0.09087893467819544
diff --git a/logs/eval_NPO_lr1e-05_beta0.1_alpha2_epoch10.log b/logs/eval_NPO_lr1e-05_beta0.1_alpha2_epoch10.log
new file mode 100644
index 0000000000000000000000000000000000000000..148db09cb56387b9c23bc9936cf02a18c7c53a36
--- /dev/null
+++ b/logs/eval_NPO_lr1e-05_beta0.1_alpha2_epoch10.log
@@ -0,0 +1,78 @@
+/opt/conda/envs/copygen/lib/python3.12/site-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.6.3) or chardet (6.0.0.post1)/charset_normalizer (3.4.4) doesn't match a supported version!
+  warnings.warn(
+[2026-07-22 17:20:39,331][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/NPO_lr1e-05_beta0.1_alpha2_epoch10
+[2026-07-22 17:20:39,332][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-22 17:20:39,332][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/NPO_lr1e-05_beta0.1_alpha2_epoch10/TOFU_EVAL.json
+[2026-07-22 17:20:39,332][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/NPO_lr1e-05_beta0.1_alpha2_epoch10/TOFU_SUMMARY.json
+[2026-07-22 17:20:39,979][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:05,  2.08it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:03,  3.02it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  3.46it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.75it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:02,  3.99it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.12it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.23it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:02<00:01,  4.28it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.34it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.40it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.38it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.37it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.57it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.09it/s]
+Evaluated 400 examples
+[2026-07-22 17:20:43,524][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:13,  1.12s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:12,  1.13s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:11,  1.16s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:10,  1.15s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:09,  1.14s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.13s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.13s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:09<00:05,  1.11s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:10<00:04,  1.11s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:11<00:03,  1.11s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:12<00:02,  1.12s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.12s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.07s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.11s/it]
+Evaluated 400 examples
+[2026-07-22 17:20:57,974][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-22 17:20:57,974][evaluator][INFO] - Result for metric forget_truth_ratio:	0.5920935820314592
+[2026-07-22 17:20:57,978][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-22 17:20:57,979][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:20:57,986][metrics][INFO] - Evaluating forget_quality
+[2026-07-22 17:20:57,987][evaluator][INFO] - Result for metric forget_quality:	9.074891594369707e-08
+[2026-07-22 17:21:03,369][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.40it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.44it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.30it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:02,  4.33it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.37it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.42it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.42it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.51it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.55it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.53it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.51it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.49it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.73it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.52it/s]
+Evaluated 400 examples
+[2026-07-22 17:21:06,247][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.2562120056152344
+[2026-07-22 17:21:06,524][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:02,  4.09it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:03,  3.52it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:02,  3.61it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:02,  3.58it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:02,  3.68it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:01<00:01,  3.66it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:01<00:01,  3.75it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  3.69it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:02<00:01,  3.65it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:02<00:00,  3.37it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:03<00:00,  3.66it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:03<00:00,  3.66it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.79it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.67it/s]
+Evaluated 400 examples
+[2026-07-22 17:21:10,063][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.21079787542832967
+[2026-07-22 17:21:10,381][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.56it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.65it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.68it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.70it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.68it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.58it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.56it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.52it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.57it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.59it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.55it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.61it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.84it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.66it/s]
+Evaluated 400 examples
+[2026-07-22 17:21:13,541][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:02,  4.26it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:02,  3.87it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:02,  4.32it/s]
Calculating text similarity:  31%|███       | 4/13 [00:00<00:02,  4.29it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:02,  3.83it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:01<00:01,  3.66it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:01<00:01,  3.80it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  3.97it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:02<00:00,  4.03it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:02<00:00,  3.93it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:02<00:00,  3.57it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:03<00:00,  3.60it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  2.84it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.55it/s]
+Evaluated 400 examples
+[2026-07-22 17:21:17,454][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.59it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.60it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.57it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.63it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.65it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.57it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.53it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.48it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.48it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.44it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.39it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.42it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.66it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.55it/s]
+Evaluated 400 examples
+[2026-07-22 17:21:20,628][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:12,  1.06s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:11,  1.06s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:10,  1.07s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:09,  1.07s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:08,  1.06s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.08s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.10s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.10s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:09<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:10<00:03,  1.10s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:11<00:02,  1.11s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.10s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.05s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.08s/it]
+Evaluated 400 examples
+[2026-07-22 17:21:34,642][metrics][INFO] - Evaluating retain_Truth_Ratio
+[2026-07-22 17:21:35,213][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.11it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.52it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  7.52it/s]
+Evaluated 100 examples
+[2026-07-22 17:21:35,988][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.28it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.45it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.08it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.48it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.51it/s]
+Evaluated 100 examples
+[2026-07-22 17:21:37,583][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+[2026-07-22 17:21:37,895][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  7.00it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  5.78it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  5.67it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  6.60it/s]
+Evaluated 100 examples
+[2026-07-22 17:21:38,502][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-22 17:21:38,502][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:21:38,502][metrics][INFO] - Evaluating ra_Truth_Ratio
+[2026-07-22 17:21:38,740][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.17it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.64it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  8.43it/s]
+Evaluated 117 examples
+[2026-07-22 17:21:39,520][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.83it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.51it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.09it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.55it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.57it/s]
+Evaluated 117 examples
+[2026-07-22 17:21:41,076][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+[2026-07-22 17:21:41,314][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  5.57it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  5.48it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  5.35it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  4.32it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  4.67it/s]
+Evaluated 117 examples
+[2026-07-22 17:21:42,172][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-22 17:21:42,172][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:21:42,172][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-22 17:21:42,173][metrics][INFO] - Evaluating model_utility
+[2026-07-22 17:21:42,173][evaluator][INFO] - Result for metric model_utility:	0.4934882075014201
+[2026-07-22 17:21:42,667][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:21:42,678][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.68it/s]
 31%|███       | 4/13 [00:00<00:00, 12.40it/s]
 46%|████▌     | 6/13 [00:00<00:00, 13.01it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.07it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.54it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.23it/s]
100%|██████████| 13/13 [00:00<00:00, 13.73it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.16it/s]
 31%|███       | 4/13 [00:00<00:00, 12.74it/s]
 46%|████▌     | 6/13 [00:00<00:00, 12.93it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 12.97it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.07it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.17it/s]
100%|██████████| 13/13 [00:00<00:00, 13.50it/s]
+[2026-07-22 17:21:44,590][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:21:44,598][metrics][INFO] - Evaluating privleak
+[2026-07-22 17:21:44,598][evaluator][INFO] - Result for metric privleak:	-73.8494867366363
+[2026-07-22 17:21:44,859][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:00, 11.64it/s]
Calculating ES:  31%|███       | 4/13 [00:00<00:00, 10.66it/s]
Calculating ES:  46%|████▌     | 6/13 [00:00<00:00, 11.28it/s]
Calculating ES:  62%|██████▏   | 8/13 [00:00<00:00, 11.71it/s]
Calculating ES:  77%|███████▋  | 10/13 [00:00<00:00, 11.91it/s]
Calculating ES:  92%|█████████▏| 12/13 [00:01<00:00, 11.73it/s]
Calculating ES: 100%|██████████| 13/13 [00:01<00:00, 12.09it/s]
+Evaluated 400 examples
+[2026-07-22 17:21:45,935][evaluator][INFO] - Result for metric extraction_strength:	0.09733772122524723
diff --git a/logs/eval_NPO_lr1e-05_beta0.1_alpha2_epoch5.log b/logs/eval_NPO_lr1e-05_beta0.1_alpha2_epoch5.log
new file mode 100644
index 0000000000000000000000000000000000000000..89b5656a30ab5bb7615ff7710faee3a15287a29f
--- /dev/null
+++ b/logs/eval_NPO_lr1e-05_beta0.1_alpha2_epoch5.log
@@ -0,0 +1,78 @@
+/opt/conda/envs/copygen/lib/python3.12/site-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.6.3) or chardet (6.0.0.post1)/charset_normalizer (3.4.4) doesn't match a supported version!
+  warnings.warn(
+[2026-07-22 17:28:29,682][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/NPO_lr1e-05_beta0.1_alpha2_epoch5
+[2026-07-22 17:28:29,683][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-22 17:28:29,683][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/NPO_lr1e-05_beta0.1_alpha2_epoch5/TOFU_EVAL.json
+[2026-07-22 17:28:29,683][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/NPO_lr1e-05_beta0.1_alpha2_epoch5/TOFU_SUMMARY.json
+[2026-07-22 17:28:30,509][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:05,  2.30it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:03,  3.20it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  3.58it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.82it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.03it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.10it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.21it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:02<00:01,  4.34it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.34it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.36it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.32it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.31it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.52it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.11it/s]
+Evaluated 400 examples
+[2026-07-22 17:28:34,076][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:13,  1.12s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:12,  1.12s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:11,  1.15s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:10,  1.15s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:09,  1.13s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.13s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.13s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:09<00:05,  1.11s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:10<00:04,  1.11s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:11<00:03,  1.10s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:12<00:02,  1.12s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.12s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.06s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.11s/it]
+Evaluated 400 examples
+[2026-07-22 17:28:48,480][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-22 17:28:48,481][evaluator][INFO] - Result for metric forget_truth_ratio:	0.6604126390365841
+[2026-07-22 17:28:48,487][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-22 17:28:48,487][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:28:48,497][metrics][INFO] - Evaluating forget_quality
+[2026-07-22 17:28:48,498][evaluator][INFO] - Result for metric forget_quality:	3.596289960343452e-05
+[2026-07-22 17:28:48,925][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.45it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.43it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.21it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:02,  4.24it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.28it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.35it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.38it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.49it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.54it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.54it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.52it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.50it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.74it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.49it/s]
+Evaluated 400 examples
+[2026-07-22 17:28:51,819][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.29440185546875
+[2026-07-22 17:28:52,133][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:02,  4.24it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:02,  4.15it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:02,  4.08it/s]
Calculating text similarity:  31%|███       | 4/13 [00:00<00:02,  4.10it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:01,  4.27it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:01<00:01,  4.27it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:01<00:01,  4.19it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:01<00:01,  4.46it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:02<00:00,  4.28it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:02<00:00,  4.00it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:02<00:00,  4.00it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:02<00:00,  4.02it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  4.23it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  4.18it/s]
+Evaluated 400 examples
+[2026-07-22 17:28:55,244][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.17512879147411575
+[2026-07-22 17:28:55,605][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.62it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.68it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.70it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.71it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.72it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.65it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.64it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.60it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.61it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.61it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.52it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.52it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.73it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.65it/s]
+Evaluated 400 examples
+[2026-07-22 17:28:58,818][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:03,  3.69it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:02,  4.08it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:02,  4.51it/s]
Calculating text similarity:  31%|███       | 4/13 [00:00<00:01,  4.60it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:01,  4.54it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:01<00:01,  4.76it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:01<00:01,  4.95it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:01<00:01,  4.61it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:01<00:00,  4.52it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:02<00:00,  4.47it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:02<00:00,  4.52it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:02<00:00,  4.63it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:02<00:00,  5.05it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:02<00:00,  4.66it/s]
+Evaluated 400 examples
+[2026-07-22 17:29:01,934][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.62it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.61it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.51it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.59it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.63it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.56it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.52it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.48it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.48it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.45it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.42it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.46it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.71it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.56it/s]
+Evaluated 400 examples
+[2026-07-22 17:29:05,094][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:13,  1.09s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:11,  1.08s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:10,  1.08s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:09,  1.07s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:08,  1.06s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.08s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.09s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.09s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:09<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:10<00:03,  1.10s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:11<00:02,  1.10s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.10s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.05s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.08s/it]
+Evaluated 400 examples
+[2026-07-22 17:29:19,112][metrics][INFO] - Evaluating retain_Truth_Ratio
+[2026-07-22 17:29:19,538][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.12it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.53it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  7.48it/s]
+Evaluated 100 examples
+[2026-07-22 17:29:20,378][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.25it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.48it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.08it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.45it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.50it/s]
+Evaluated 100 examples
+[2026-07-22 17:29:21,982][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+[2026-07-22 17:29:22,288][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  6.00it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  6.91it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  6.21it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  6.63it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  6.54it/s]
+Evaluated 100 examples
+[2026-07-22 17:29:22,901][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-22 17:29:22,901][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:29:22,901][metrics][INFO] - Evaluating ra_Truth_Ratio
+[2026-07-22 17:29:23,201][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.19it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.65it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  8.46it/s]
+Evaluated 117 examples
+[2026-07-22 17:29:23,993][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.76it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.50it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.10it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.58it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.59it/s]
+Evaluated 117 examples
+[2026-07-22 17:29:25,538][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+[2026-07-22 17:29:25,845][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  7.79it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  6.94it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  6.96it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  6.32it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  6.59it/s]
+Evaluated 117 examples
+[2026-07-22 17:29:26,453][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-22 17:29:26,454][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:29:26,454][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-22 17:29:26,454][metrics][INFO] - Evaluating model_utility
+[2026-07-22 17:29:26,454][evaluator][INFO] - Result for metric model_utility:	0.4182503538554803
+[2026-07-22 17:29:27,114][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:29:27,122][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.96it/s]
 31%|███       | 4/13 [00:00<00:00, 12.53it/s]
 46%|████▌     | 6/13 [00:00<00:00, 13.26it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.65it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.96it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.71it/s]
100%|██████████| 13/13 [00:00<00:00, 14.16it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.43it/s]
 31%|███       | 4/13 [00:00<00:00, 13.35it/s]
 46%|████▌     | 6/13 [00:00<00:00, 13.75it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.73it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.82it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.72it/s]
100%|██████████| 13/13 [00:00<00:00, 14.15it/s]
+[2026-07-22 17:29:28,960][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:29:28,968][metrics][INFO] - Evaluating privleak
+[2026-07-22 17:29:28,968][evaluator][INFO] - Result for metric privleak:	-79.42388288078007
+[2026-07-22 17:29:29,337][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:01, 10.73it/s]
Calculating ES:  31%|███       | 4/13 [00:00<00:00,  9.82it/s]
Calculating ES:  46%|████▌     | 6/13 [00:00<00:00, 10.49it/s]
Calculating ES:  62%|██████▏   | 8/13 [00:00<00:00, 10.87it/s]
Calculating ES:  77%|███████▋  | 10/13 [00:00<00:00, 11.03it/s]
Calculating ES:  92%|█████████▏| 12/13 [00:01<00:00, 10.74it/s]
Calculating ES: 100%|██████████| 13/13 [00:01<00:00, 11.14it/s]
+Evaluated 400 examples
+[2026-07-22 17:29:30,505][evaluator][INFO] - Result for metric extraction_strength:	0.10077445856134101
diff --git a/logs/eval_NPO_lr1e-05_beta0.1_alpha5_epoch10.log b/logs/eval_NPO_lr1e-05_beta0.1_alpha5_epoch10.log
new file mode 100644
index 0000000000000000000000000000000000000000..ae803044117c053646b6082eb05d880df21bb191
--- /dev/null
+++ b/logs/eval_NPO_lr1e-05_beta0.1_alpha5_epoch10.log
@@ -0,0 +1,124 @@
+[2026-07-21 07:51:50,231][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/NPO_lr1e-05_beta0.1_alpha5_epoch10
+[2026-07-21 07:51:50,232][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-21 07:51:50,232][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/NPO_lr1e-05_beta0.1_alpha5_epoch10/TOFU_EVAL.json
+[2026-07-21 07:51:50,232][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/NPO_lr1e-05_beta0.1_alpha5_epoch10/TOFU_SUMMARY.json
+[2026-07-21 07:51:50,927][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:07,  1.50it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:04,  2.51it/s]
Calculating loss:  23%|██▎       | 3/13 [00:01<00:03,  3.15it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.61it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:02,  3.96it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.18it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.37it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:02<00:01,  4.54it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.62it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.67it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.68it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.68it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.83it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.09it/s]
+Evaluated 400 examples
+[2026-07-21 07:51:54,463][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:12,  1.03s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:11,  1.02s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:10,  1.05s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:09,  1.05s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:08,  1.04s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.03s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.03s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.01s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:09<00:04,  1.01s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:10<00:03,  1.01s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:11<00:02,  1.02s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:12<00:01,  1.02s/it]
Calculating loss: 100%|██████████| 13/13 [00:13<00:00,  1.03it/s]
Calculating loss: 100%|██████████| 13/13 [00:13<00:00,  1.01s/it]
+Evaluated 400 examples
+[2026-07-21 07:52:07,618][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-21 07:52:07,619][evaluator][INFO] - Result for metric forget_truth_ratio:	0.6084976365334408
+[2026-07-21 07:52:07,627][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-21 07:52:07,628][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-21 07:52:07,644][metrics][INFO] - Evaluating forget_quality
+[2026-07-21 07:52:07,645][evaluator][INFO] - Result for metric forget_quality:	4.455792684961036e-06
+[2026-07-21 07:52:08,041][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.80it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.81it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.67it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.70it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.72it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.77it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.77it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.85it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.88it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.86it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.84it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.82it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  5.08it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.86it/s]
+Evaluated 400 examples
+[2026-07-21 07:52:10,715][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.3352020263671875
+[2026-07-21 07:52:10,993][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:07,  1.70it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:01<00:07,  1.45it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:02<00:06,  1.46it/s]
Calculating text similarity:  31%|███       | 4/13 [00:02<00:05,  1.54it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:03<00:05,  1.49it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:03<00:04,  1.52it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:04<00:03,  1.59it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:04<00:02,  1.75it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:05<00:02,  1.80it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:06<00:01,  1.84it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:06<00:01,  1.89it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:07<00:00,  1.80it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:07<00:00,  1.69it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:07<00:00,  1.67it/s]
+Evaluated 400 examples
+[2026-07-21 07:52:18,801][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.3103877653460595
+[2026-07-21 07:52:19,154][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.94it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.98it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:01,  5.01it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  5.03it/s]
Calculating loss:  38%|███▊      | 5/13 [00:00<00:01,  5.04it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.97it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.96it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.92it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.95it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.94it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.89it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.93it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  5.17it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  5.01it/s]
+Using the latest cached version of the dataset since locuslab/TOFU couldn't be found on the Hugging Face Hub
+Evaluated 400 examples
+[2026-07-21 07:52:21,786][datasets.load][WARNING] - Using the latest cached version of the dataset since locuslab/TOFU couldn't be found on the Hugging Face Hub
+Found the latest cached dataset configuration 'retain_perturbed' at /workspace/hf_cache_ou/locuslab___tofu/retain_perturbed/0.0.0/324592d84ae4f482ac7249b9285c2ecdb53e3a68 (last modified on Mon Jul 20 13:22:39 2026).
+[2026-07-21 07:52:21,787][datasets.packaged_modules.cache.cache][WARNING] - Found the latest cached dataset configuration 'retain_perturbed' at /workspace/hf_cache_ou/locuslab___tofu/retain_perturbed/0.0.0/324592d84ae4f482ac7249b9285c2ecdb53e3a68 (last modified on Mon Jul 20 13:22:39 2026).
+[2026-07-21 07:52:21,794][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:05,  2.09it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:01<00:06,  1.81it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:01<00:05,  1.74it/s]
Calculating text similarity:  31%|███       | 4/13 [00:02<00:04,  1.81it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:02<00:04,  1.90it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:03<00:03,  1.88it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:03<00:03,  1.79it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:04<00:02,  1.82it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:04<00:02,  1.77it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:05<00:01,  1.82it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:06<00:01,  1.67it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:06<00:00,  1.71it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:07<00:00,  1.79it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:07<00:00,  1.79it/s]
+Using the latest cached version of the dataset since locuslab/TOFU couldn't be found on the Hugging Face Hub
+Evaluated 400 examples
+[2026-07-21 07:52:29,082][datasets.load][WARNING] - Using the latest cached version of the dataset since locuslab/TOFU couldn't be found on the Hugging Face Hub
+Found the latest cached dataset configuration 'retain_perturbed' at /workspace/hf_cache_ou/locuslab___tofu/retain_perturbed/0.0.0/324592d84ae4f482ac7249b9285c2ecdb53e3a68 (last modified on Mon Jul 20 13:22:39 2026).
+[2026-07-21 07:52:29,084][datasets.packaged_modules.cache.cache][WARNING] - Found the latest cached dataset configuration 'retain_perturbed' at /workspace/hf_cache_ou/locuslab___tofu/retain_perturbed/0.0.0/324592d84ae4f482ac7249b9285c2ecdb53e3a68 (last modified on Mon Jul 20 13:22:39 2026).
+[2026-07-21 07:52:29,089][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.92it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.94it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.91it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.96it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.99it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.91it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.88it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.82it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.82it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.82it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.80it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.82it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  5.08it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.92it/s]
+Using the latest cached version of the dataset since locuslab/TOFU couldn't be found on the Hugging Face Hub
+Evaluated 400 examples
+[2026-07-21 07:52:31,768][datasets.load][WARNING] - Using the latest cached version of the dataset since locuslab/TOFU couldn't be found on the Hugging Face Hub
+Found the latest cached dataset configuration 'retain_perturbed' at /workspace/hf_cache_ou/locuslab___tofu/retain_perturbed/0.0.0/324592d84ae4f482ac7249b9285c2ecdb53e3a68 (last modified on Mon Jul 20 13:22:39 2026).
+[2026-07-21 07:52:31,771][datasets.packaged_modules.cache.cache][WARNING] - Found the latest cached dataset configuration 'retain_perturbed' at /workspace/hf_cache_ou/locuslab___tofu/retain_perturbed/0.0.0/324592d84ae4f482ac7249b9285c2ecdb53e3a68 (last modified on Mon Jul 20 13:22:39 2026).
+[2026-07-21 07:52:31,775][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:11,  1.02it/s]
Calculating loss:  15%|█▌        | 2/13 [00:01<00:10,  1.02it/s]
Calculating loss:  23%|██▎       | 3/13 [00:02<00:09,  1.01it/s]
Calculating loss:  31%|███       | 4/13 [00:03<00:08,  1.02it/s]
Calculating loss:  38%|███▊      | 5/13 [00:04<00:07,  1.02it/s]
Calculating loss:  46%|████▌     | 6/13 [00:05<00:06,  1.01it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:06<00:05,  1.00it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:07<00:05,  1.00s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:08<00:04,  1.01s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:09<00:03,  1.01s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:10<00:02,  1.01s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:11<00:01,  1.01s/it]
Calculating loss: 100%|██████████| 13/13 [00:12<00:00,  1.04it/s]
Calculating loss: 100%|██████████| 13/13 [00:12<00:00,  1.01it/s]
+Evaluated 400 examples
+[2026-07-21 07:52:44,621][metrics][INFO] - Evaluating retain_Truth_Ratio
+Using the latest cached version of the dataset since locuslab/TOFU couldn't be found on the Hugging Face Hub
+[2026-07-21 07:52:44,661][datasets.load][WARNING] - Using the latest cached version of the dataset since locuslab/TOFU couldn't be found on the Hugging Face Hub
+Found the latest cached dataset configuration 'real_authors_perturbed' at /workspace/hf_cache_ou/locuslab___tofu/real_authors_perturbed/0.0.0/324592d84ae4f482ac7249b9285c2ecdb53e3a68 (last modified on Mon Jul 20 13:23:07 2026).
+[2026-07-21 07:52:44,662][datasets.packaged_modules.cache.cache][WARNING] - Found the latest cached dataset configuration 'real_authors_perturbed' at /workspace/hf_cache_ou/locuslab___tofu/real_authors_perturbed/0.0.0/324592d84ae4f482ac7249b9285c2ecdb53e3a68 (last modified on Mon Jul 20 13:23:07 2026).
+[2026-07-21 07:52:44,668][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.44it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.86it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  8.00it/s]
+Using the latest cached version of the dataset since locuslab/TOFU couldn't be found on the Hugging Face Hub
+Evaluated 100 examples
+[2026-07-21 07:52:45,204][datasets.load][WARNING] - Using the latest cached version of the dataset since locuslab/TOFU couldn't be found on the Hugging Face Hub
+Found the latest cached dataset configuration 'real_authors_perturbed' at /workspace/hf_cache_ou/locuslab___tofu/real_authors_perturbed/0.0.0/324592d84ae4f482ac7249b9285c2ecdb53e3a68 (last modified on Mon Jul 20 13:23:07 2026).
+[2026-07-21 07:52:45,206][datasets.packaged_modules.cache.cache][WARNING] - Found the latest cached dataset configuration 'real_authors_perturbed' at /workspace/hf_cache_ou/locuslab___tofu/real_authors_perturbed/0.0.0/324592d84ae4f482ac7249b9285c2ecdb53e3a68 (last modified on Mon Jul 20 13:23:07 2026).
+[2026-07-21 07:52:45,210][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.29it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.62it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.21it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.63it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.67it/s]
+Evaluated 100 examples
+[2026-07-21 07:52:46,711][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+Using the latest cached version of the dataset since locuslab/TOFU couldn't be found on the Hugging Face Hub
+[2026-07-21 07:52:46,750][datasets.load][WARNING] - Using the latest cached version of the dataset since locuslab/TOFU couldn't be found on the Hugging Face Hub
+Found the latest cached dataset configuration 'real_authors_perturbed' at /workspace/hf_cache_ou/locuslab___tofu/real_authors_perturbed/0.0.0/324592d84ae4f482ac7249b9285c2ecdb53e3a68 (last modified on Mon Jul 20 13:23:07 2026).
+[2026-07-21 07:52:46,752][datasets.packaged_modules.cache.cache][WARNING] - Found the latest cached dataset configuration 'real_authors_perturbed' at /workspace/hf_cache_ou/locuslab___tofu/real_authors_perturbed/0.0.0/324592d84ae4f482ac7249b9285c2ecdb53e3a68 (last modified on Mon Jul 20 13:23:07 2026).
+[2026-07-21 07:52:46,756][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:01,  2.97it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  3.14it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  3.11it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  3.54it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  3.36it/s]
+Evaluated 100 examples
+[2026-07-21 07:52:47,947][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-21 07:52:47,947][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-21 07:52:47,947][metrics][INFO] - Evaluating ra_Truth_Ratio
+Using the latest cached version of the dataset since locuslab/TOFU couldn't be found on the Hugging Face Hub
+[2026-07-21 07:52:47,983][datasets.load][WARNING] - Using the latest cached version of the dataset since locuslab/TOFU couldn't be found on the Hugging Face Hub
+Found the latest cached dataset configuration 'world_facts_perturbed' at /workspace/hf_cache_ou/locuslab___tofu/world_facts_perturbed/0.0.0/324592d84ae4f482ac7249b9285c2ecdb53e3a68 (last modified on Mon Jul 20 13:23:12 2026).
+[2026-07-21 07:52:47,985][datasets.packaged_modules.cache.cache][WARNING] - Found the latest cached dataset configuration 'world_facts_perturbed' at /workspace/hf_cache_ou/locuslab___tofu/world_facts_perturbed/0.0.0/324592d84ae4f482ac7249b9285c2ecdb53e3a68 (last modified on Mon Jul 20 13:23:12 2026).
+[2026-07-21 07:52:47,989][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.30it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.83it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  8.56it/s]
+Using the latest cached version of the dataset since locuslab/TOFU couldn't be found on the Hugging Face Hub
+Evaluated 117 examples
+[2026-07-21 07:52:48,495][datasets.load][WARNING] - Using the latest cached version of the dataset since locuslab/TOFU couldn't be found on the Hugging Face Hub
+Found the latest cached dataset configuration 'world_facts_perturbed' at /workspace/hf_cache_ou/locuslab___tofu/world_facts_perturbed/0.0.0/324592d84ae4f482ac7249b9285c2ecdb53e3a68 (last modified on Mon Jul 20 13:23:12 2026).
+[2026-07-21 07:52:48,497][datasets.packaged_modules.cache.cache][WARNING] - Found the latest cached dataset configuration 'world_facts_perturbed' at /workspace/hf_cache_ou/locuslab___tofu/world_facts_perturbed/0.0.0/324592d84ae4f482ac7249b9285c2ecdb53e3a68 (last modified on Mon Jul 20 13:23:12 2026).
+[2026-07-21 07:52:48,501][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.47it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.61it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.21it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.70it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.71it/s]
+Evaluated 117 examples
+[2026-07-21 07:52:49,978][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+Using the latest cached version of the dataset since locuslab/TOFU couldn't be found on the Hugging Face Hub
+[2026-07-21 07:52:50,013][datasets.load][WARNING] - Using the latest cached version of the dataset since locuslab/TOFU couldn't be found on the Hugging Face Hub
+Found the latest cached dataset configuration 'world_facts_perturbed' at /workspace/hf_cache_ou/locuslab___tofu/world_facts_perturbed/0.0.0/324592d84ae4f482ac7249b9285c2ecdb53e3a68 (last modified on Mon Jul 20 13:23:12 2026).
+[2026-07-21 07:52:50,015][datasets.packaged_modules.cache.cache][WARNING] - Found the latest cached dataset configuration 'world_facts_perturbed' at /workspace/hf_cache_ou/locuslab___tofu/world_facts_perturbed/0.0.0/324592d84ae4f482ac7249b9285c2ecdb53e3a68 (last modified on Mon Jul 20 13:23:12 2026).
+[2026-07-21 07:52:50,020][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:01,  2.40it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  2.36it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:01<00:00,  2.35it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  2.55it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  2.48it/s]
+Evaluated 117 examples
+[2026-07-21 07:52:51,636][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-21 07:52:51,636][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-21 07:52:51,636][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-21 07:52:51,637][metrics][INFO] - Evaluating model_utility
+[2026-07-21 07:52:51,638][evaluator][INFO] - Result for metric model_utility:	0.5583081516347986
+Using the latest cached version of the dataset since locuslab/TOFU couldn't be found on the Hugging Face Hub
+[2026-07-21 07:52:51,707][datasets.load][WARNING] - Using the latest cached version of the dataset since locuslab/TOFU couldn't be found on the Hugging Face Hub
+Found the latest cached dataset configuration 'forget10_perturbed' at /workspace/hf_cache_ou/locuslab___tofu/forget10_perturbed/0.0.0/324592d84ae4f482ac7249b9285c2ecdb53e3a68 (last modified on Mon Jul 20 13:17:15 2026).
+[2026-07-21 07:52:51,709][datasets.packaged_modules.cache.cache][WARNING] - Found the latest cached dataset configuration 'forget10_perturbed' at /workspace/hf_cache_ou/locuslab___tofu/forget10_perturbed/0.0.0/324592d84ae4f482ac7249b9285c2ecdb53e3a68 (last modified on Mon Jul 20 13:17:15 2026).
+Using the latest cached version of the dataset since locuslab/TOFU couldn't be found on the Hugging Face Hub
+[2026-07-21 07:52:51,749][datasets.load][WARNING] - Using the latest cached version of the dataset since locuslab/TOFU couldn't be found on the Hugging Face Hub
+Found the latest cached dataset configuration 'holdout10' at /workspace/hf_cache_ou/locuslab___tofu/holdout10/0.0.0/324592d84ae4f482ac7249b9285c2ecdb53e3a68 (last modified on Mon Jul 20 13:23:17 2026).
+[2026-07-21 07:52:51,752][datasets.packaged_modules.cache.cache][WARNING] - Found the latest cached dataset configuration 'holdout10' at /workspace/hf_cache_ou/locuslab___tofu/holdout10/0.0.0/324592d84ae4f482ac7249b9285c2ecdb53e3a68 (last modified on Mon Jul 20 13:23:17 2026).
+[2026-07-21 07:52:51,756][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-21 07:52:51,771][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.22it/s]
 31%|███       | 4/13 [00:00<00:00, 11.29it/s]
 46%|████▌     | 6/13 [00:00<00:00, 11.58it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 11.72it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 11.97it/s]
 92%|█████████▏| 12/13 [00:01<00:00, 11.73it/s]
100%|██████████| 13/13 [00:01<00:00, 12.27it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 11.45it/s]
 31%|███       | 4/13 [00:00<00:00, 11.30it/s]
 46%|████▌     | 6/13 [00:00<00:00, 11.60it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 11.67it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 11.70it/s]
 92%|█████████▏| 12/13 [00:01<00:00, 11.83it/s]
100%|██████████| 13/13 [00:01<00:00, 12.10it/s]
+[2026-07-21 07:52:53,910][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-21 07:52:53,924][metrics][INFO] - Evaluating privleak
+[2026-07-21 07:52:53,924][evaluator][INFO] - Result for metric privleak:	-80.60122288180416
+Using the latest cached version of the dataset since locuslab/TOFU couldn't be found on the Hugging Face Hub
+[2026-07-21 07:52:53,996][datasets.load][WARNING] - Using the latest cached version of the dataset since locuslab/TOFU couldn't be found on the Hugging Face Hub
+Found the latest cached dataset configuration 'forget10_perturbed' at /workspace/hf_cache_ou/locuslab___tofu/forget10_perturbed/0.0.0/324592d84ae4f482ac7249b9285c2ecdb53e3a68 (last modified on Mon Jul 20 13:17:15 2026).
+[2026-07-21 07:52:53,998][datasets.packaged_modules.cache.cache][WARNING] - Found the latest cached dataset configuration 'forget10_perturbed' at /workspace/hf_cache_ou/locuslab___tofu/forget10_perturbed/0.0.0/324592d84ae4f482ac7249b9285c2ecdb53e3a68 (last modified on Mon Jul 20 13:17:15 2026).
+[2026-07-21 07:52:54,002][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:01, 10.15it/s]
Calculating ES:  31%|███       | 4/13 [00:00<00:00,  9.16it/s]
Calculating ES:  38%|███▊      | 5/13 [00:00<00:00,  9.36it/s]
Calculating ES:  54%|█████▍    | 7/13 [00:00<00:00,  9.64it/s]
Calculating ES:  69%|██████▉   | 9/13 [00:00<00:00, 10.07it/s]
Calculating ES:  85%|████████▍ | 11/13 [00:01<00:00,  9.92it/s]
Calculating ES:  92%|█████████▏| 12/13 [00:01<00:00,  9.78it/s]
Calculating ES: 100%|██████████| 13/13 [00:01<00:00, 10.12it/s]
+Evaluated 400 examples
+[2026-07-21 07:52:55,288][evaluator][INFO] - Result for metric extraction_strength:	0.10607682881132786
diff --git a/logs/eval_NPO_lr1e-05_beta0.1_alpha5_epoch5.log b/logs/eval_NPO_lr1e-05_beta0.1_alpha5_epoch5.log
new file mode 100644
index 0000000000000000000000000000000000000000..a8de6812df6ab39e8e4bc178eb13442d43c42a34
--- /dev/null
+++ b/logs/eval_NPO_lr1e-05_beta0.1_alpha5_epoch5.log
@@ -0,0 +1,78 @@
+/opt/conda/envs/copygen/lib/python3.12/site-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.6.3) or chardet (6.0.0.post1)/charset_normalizer (3.4.4) doesn't match a supported version!
+  warnings.warn(
+[2026-07-22 17:22:12,001][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/NPO_lr1e-05_beta0.1_alpha5_epoch5
+[2026-07-22 17:22:12,002][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-22 17:22:12,002][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/NPO_lr1e-05_beta0.1_alpha5_epoch5/TOFU_EVAL.json
+[2026-07-22 17:22:12,002][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/NPO_lr1e-05_beta0.1_alpha5_epoch5/TOFU_SUMMARY.json
+[2026-07-22 17:22:12,723][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:05,  2.21it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:03,  3.13it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  3.50it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.76it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:02,  3.98it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.10it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.20it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:02<00:01,  4.33it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.37it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.41it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.39it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.37it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.57it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.11it/s]
+Evaluated 400 examples
+[2026-07-22 17:22:16,256][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:13,  1.13s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:12,  1.13s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:11,  1.15s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:10,  1.15s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:09,  1.13s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.13s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.13s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:09<00:05,  1.11s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:10<00:04,  1.11s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:11<00:03,  1.11s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:12<00:02,  1.12s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.12s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.07s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.11s/it]
+Evaluated 400 examples
+[2026-07-22 17:22:30,696][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-22 17:22:30,696][evaluator][INFO] - Result for metric forget_truth_ratio:	0.6256303136974539
+[2026-07-22 17:22:30,701][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-22 17:22:30,702][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:22:30,710][metrics][INFO] - Evaluating forget_quality
+[2026-07-22 17:22:30,711][evaluator][INFO] - Result for metric forget_quality:	3.081974946564706e-07
+[2026-07-22 17:22:31,195][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.42it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.42it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.27it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:02,  4.31it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.34it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.39it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.39it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.48it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.52it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.50it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.47it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.44it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.68it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.48it/s]
+Evaluated 400 examples
+[2026-07-22 17:22:34,096][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.3873016357421875
+[2026-07-22 17:22:34,358][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:04,  2.72it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:04,  2.72it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:01<00:03,  2.65it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:03,  2.80it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:02,  3.18it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:01<00:02,  3.32it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:02<00:01,  3.43it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  3.64it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:02<00:01,  3.70it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:03<00:00,  3.68it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:03<00:00,  3.57it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:03<00:00,  3.61it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.83it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.42it/s]
+Evaluated 400 examples
+[2026-07-22 17:22:38,164][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.3212421837095092
+[2026-07-22 17:22:38,457][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.62it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.69it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.66it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.69it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.67it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.61it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.61it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.59it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.62it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.62it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.56it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.60it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.82it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.67it/s]
+Evaluated 400 examples
+[2026-07-22 17:22:41,559][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:03,  3.51it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:03,  3.24it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:02,  3.38it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:02,  3.49it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:02,  3.57it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:01<00:01,  3.52it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:02<00:01,  3.42it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  3.46it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:02<00:01,  3.47it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:02<00:00,  3.67it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:03<00:00,  3.72it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:03<00:00,  3.84it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  4.00it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.65it/s]
+Evaluated 400 examples
+[2026-07-22 17:22:45,373][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.62it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.64it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.60it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.65it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.67it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.60it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.56it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.51it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.47it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.49it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.45it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.48it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.72it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.59it/s]
+Evaluated 400 examples
+[2026-07-22 17:22:48,467][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:12,  1.07s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:11,  1.06s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:10,  1.09s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:09,  1.07s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:08,  1.06s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.07s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.09s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.09s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:09<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:10<00:03,  1.10s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:12<00:02,  1.11s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.10s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.05s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.08s/it]
+Evaluated 400 examples
+[2026-07-22 17:23:02,489][metrics][INFO] - Evaluating retain_Truth_Ratio
+[2026-07-22 17:23:03,060][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.16it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.57it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  7.57it/s]
+Evaluated 100 examples
+[2026-07-22 17:23:03,836][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.41it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.49it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.09it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.47it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.51it/s]
+Evaluated 100 examples
+[2026-07-22 17:23:05,431][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+[2026-07-22 17:23:05,692][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  6.28it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  6.12it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  5.85it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  6.74it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  6.46it/s]
+Evaluated 100 examples
+[2026-07-22 17:23:06,312][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-22 17:23:06,312][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:23:06,312][metrics][INFO] - Evaluating ra_Truth_Ratio
+[2026-07-22 17:23:06,585][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  7.99it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.54it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  8.21it/s]
+Evaluated 117 examples
+[2026-07-22 17:23:07,324][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.46it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.42it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.07it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.55it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.55it/s]
+Evaluated 117 examples
+[2026-07-22 17:23:08,894][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+[2026-07-22 17:23:09,175][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  5.21it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  5.67it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  3.94it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  4.50it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  4.56it/s]
+Evaluated 117 examples
+[2026-07-22 17:23:10,052][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-22 17:23:10,053][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:23:10,053][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-22 17:23:10,053][metrics][INFO] - Evaluating model_utility
+[2026-07-22 17:23:10,053][evaluator][INFO] - Result for metric model_utility:	0.5301684588487601
+[2026-07-22 17:23:10,604][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:23:10,611][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 14.07it/s]
 31%|███       | 4/13 [00:00<00:00, 12.69it/s]
 46%|████▌     | 6/13 [00:00<00:00, 13.28it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.61it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.94it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.65it/s]
100%|██████████| 13/13 [00:00<00:00, 14.15it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.12it/s]
 31%|███       | 4/13 [00:00<00:00, 12.95it/s]
 46%|████▌     | 6/13 [00:00<00:00, 13.11it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.20it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.28it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.05it/s]
100%|██████████| 13/13 [00:00<00:00, 13.50it/s]
+[2026-07-22 17:23:12,495][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:23:12,502][metrics][INFO] - Evaluating privleak
+[2026-07-22 17:23:12,502][evaluator][INFO] - Result for metric privleak:	-84.95778581766594
+[2026-07-22 17:23:12,780][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:00, 11.31it/s]
Calculating ES:  31%|███       | 4/13 [00:00<00:00, 10.24it/s]
Calculating ES:  46%|████▌     | 6/13 [00:00<00:00, 10.71it/s]
Calculating ES:  62%|██████▏   | 8/13 [00:00<00:00, 11.10it/s]
Calculating ES:  77%|███████▋  | 10/13 [00:00<00:00, 11.16it/s]
Calculating ES:  92%|█████████▏| 12/13 [00:01<00:00, 10.83it/s]
Calculating ES: 100%|██████████| 13/13 [00:01<00:00, 11.29it/s]
+Evaluated 400 examples
+[2026-07-22 17:23:13,932][evaluator][INFO] - Result for metric extraction_strength:	0.11665128371321863
diff --git a/logs/eval_NPO_lr1e-05_beta0.5_alpha1_epoch10.log b/logs/eval_NPO_lr1e-05_beta0.5_alpha1_epoch10.log
new file mode 100644
index 0000000000000000000000000000000000000000..8eae38b0cb914a51f81ef6a988c616fabf1b356c
--- /dev/null
+++ b/logs/eval_NPO_lr1e-05_beta0.5_alpha1_epoch10.log
@@ -0,0 +1,78 @@
+/opt/conda/envs/copygen/lib/python3.12/site-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.6.3) or chardet (6.0.0.post1)/charset_normalizer (3.4.4) doesn't match a supported version!
+  warnings.warn(
+[2026-07-22 17:29:59,868][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/NPO_lr1e-05_beta0.5_alpha1_epoch10
+[2026-07-22 17:29:59,869][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-22 17:29:59,869][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/NPO_lr1e-05_beta0.5_alpha1_epoch10/TOFU_EVAL.json
+[2026-07-22 17:29:59,869][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/NPO_lr1e-05_beta0.5_alpha1_epoch10/TOFU_SUMMARY.json
+[2026-07-22 17:30:00,661][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:05,  2.25it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:03,  3.13it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  3.53it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.79it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:02,  3.99it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.11it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.21it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:02<00:01,  4.34it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.39it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.43it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.41it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.38it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.56it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.13it/s]
+Evaluated 400 examples
+[2026-07-22 17:30:04,328][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:13,  1.12s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:12,  1.11s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:11,  1.14s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:10,  1.15s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:09,  1.13s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.13s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.13s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:09<00:05,  1.12s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:10<00:04,  1.11s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:11<00:03,  1.10s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:12<00:02,  1.11s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.12s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.06s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.11s/it]
+Evaluated 400 examples
+[2026-07-22 17:30:18,723][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-22 17:30:18,724][evaluator][INFO] - Result for metric forget_truth_ratio:	0.5279551499312274
+[2026-07-22 17:30:18,728][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-22 17:30:18,728][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:30:18,735][metrics][INFO] - Evaluating forget_quality
+[2026-07-22 17:30:18,736][evaluator][INFO] - Result for metric forget_quality:	1.0229984068189572e-13
+[2026-07-22 17:30:19,234][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.47it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.46it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.29it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:02,  4.33it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.37it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.41it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.37it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.43it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.46it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.44it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.44it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.44it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.69it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.48it/s]
+Evaluated 400 examples
+[2026-07-22 17:30:22,139][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.5689794921875
+[2026-07-22 17:30:22,505][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:03,  3.03it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:04,  2.70it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:01<00:04,  2.36it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:03,  2.57it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:02,  2.77it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:02<00:02,  3.03it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:02<00:02,  2.92it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  3.16it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:03<00:01,  3.01it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:03<00:01,  2.84it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:03<00:00,  2.91it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:04<00:00,  2.78it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:04<00:00,  2.87it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:04<00:00,  2.85it/s]
+Evaluated 400 examples
+[2026-07-22 17:30:27,075][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.448793508187839
+[2026-07-22 17:30:27,529][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.51it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.64it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.64it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.69it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.70it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.64it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.63it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.56it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.57it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.55it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.49it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.55it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.79it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.64it/s]
+Evaluated 400 examples
+[2026-07-22 17:30:30,623][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:03,  3.12it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:03,  3.47it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:02,  3.43it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:02,  3.71it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:02,  3.65it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:01<00:01,  3.57it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:02<00:01,  3.33it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  3.26it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:02<00:01,  3.26it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:03<00:00,  3.12it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:03<00:00,  2.93it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:03<00:00,  3.10it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.39it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.32it/s]
+Evaluated 400 examples
+[2026-07-22 17:30:34,933][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.60it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.60it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.56it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.62it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.64it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.57it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.54it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.48it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.49it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.50it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.48it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.47it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.68it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.57it/s]
+Evaluated 400 examples
+[2026-07-22 17:30:38,129][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:12,  1.07s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:11,  1.06s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:10,  1.07s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:09,  1.07s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:08,  1.06s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.07s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.09s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.10s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:09<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:10<00:03,  1.09s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:11<00:02,  1.10s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.10s/it]
Calculating loss: 100%|██████████| 13/13 [00:13<00:00,  1.05s/it]
Calculating loss: 100%|██████████| 13/13 [00:13<00:00,  1.08s/it]
+Evaluated 400 examples
+[2026-07-22 17:30:52,120][metrics][INFO] - Evaluating retain_Truth_Ratio
+[2026-07-22 17:30:52,651][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.05it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.51it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  7.44it/s]
+Evaluated 100 examples
+[2026-07-22 17:30:53,510][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.03it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.43it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.05it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.44it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.47it/s]
+Evaluated 100 examples
+[2026-07-22 17:30:55,129][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+[2026-07-22 17:30:55,448][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  4.75it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  5.24it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  5.15it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  6.33it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  5.81it/s]
+Evaluated 100 examples
+[2026-07-22 17:30:56,137][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-22 17:30:56,137][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:30:56,137][metrics][INFO] - Evaluating ra_Truth_Ratio
+[2026-07-22 17:30:56,463][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.08it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.60it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  8.38it/s]
+Evaluated 117 examples
+[2026-07-22 17:30:57,262][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.91it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.51it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.10it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.57it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.59it/s]
+Evaluated 117 examples
+[2026-07-22 17:30:58,810][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+[2026-07-22 17:30:59,153][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  3.81it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  4.29it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  3.98it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  4.11it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  4.09it/s]
+Evaluated 117 examples
+[2026-07-22 17:31:00,132][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-22 17:31:00,132][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:31:00,132][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-22 17:31:00,132][metrics][INFO] - Evaluating model_utility
+[2026-07-22 17:31:00,133][evaluator][INFO] - Result for metric model_utility:	0.5935198824590285
+[2026-07-22 17:31:00,781][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:31:00,788][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.58it/s]
 31%|███       | 4/13 [00:00<00:00, 12.55it/s]
 46%|████▌     | 6/13 [00:00<00:00, 13.24it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.64it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.68it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.47it/s]
100%|██████████| 13/13 [00:00<00:00, 14.00it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 12.75it/s]
 31%|███       | 4/13 [00:00<00:00, 13.11it/s]
 46%|████▌     | 6/13 [00:00<00:00, 13.19it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.27it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.40it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.55it/s]
100%|██████████| 13/13 [00:00<00:00, 13.82it/s]
+[2026-07-22 17:31:02,661][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:31:02,668][metrics][INFO] - Evaluating privleak
+[2026-07-22 17:31:02,668][evaluator][INFO] - Result for metric privleak:	-97.25456053171555
+[2026-07-22 17:31:03,015][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:00, 11.86it/s]
Calculating ES:  31%|███       | 4/13 [00:00<00:00, 10.88it/s]
Calculating ES:  46%|████▌     | 6/13 [00:00<00:00, 11.61it/s]
Calculating ES:  62%|██████▏   | 8/13 [00:00<00:00, 12.08it/s]
Calculating ES:  77%|███████▋  | 10/13 [00:00<00:00, 12.29it/s]
Calculating ES:  92%|█████████▏| 12/13 [00:01<00:00, 12.07it/s]
Calculating ES: 100%|██████████| 13/13 [00:01<00:00, 12.43it/s]
+Evaluated 400 examples
+[2026-07-22 17:31:04,061][evaluator][INFO] - Result for metric extraction_strength:	0.18548512993763425
diff --git a/logs/eval_NPO_lr1e-05_beta0.5_alpha1_epoch5.log b/logs/eval_NPO_lr1e-05_beta0.5_alpha1_epoch5.log
new file mode 100644
index 0000000000000000000000000000000000000000..28c746c2724379a2dfa45efb8e3d90be1ae5b19b
--- /dev/null
+++ b/logs/eval_NPO_lr1e-05_beta0.5_alpha1_epoch5.log
@@ -0,0 +1,78 @@
+/opt/conda/envs/copygen/lib/python3.12/site-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.6.3) or chardet (6.0.0.post1)/charset_normalizer (3.4.4) doesn't match a supported version!
+  warnings.warn(
+[2026-07-22 17:25:50,409][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/NPO_lr1e-05_beta0.5_alpha1_epoch5
+[2026-07-22 17:25:50,410][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-22 17:25:50,410][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/NPO_lr1e-05_beta0.5_alpha1_epoch5/TOFU_EVAL.json
+[2026-07-22 17:25:50,410][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/NPO_lr1e-05_beta0.5_alpha1_epoch5/TOFU_SUMMARY.json
+[2026-07-22 17:25:51,428][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:05,  2.38it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:03,  3.26it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  3.62it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.85it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.05it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.14it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.24it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:02<00:01,  4.37it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.41it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.45it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.43it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.38it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.55it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.17it/s]
+Evaluated 400 examples
+[2026-07-22 17:25:54,977][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:13,  1.13s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:12,  1.12s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:11,  1.14s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:10,  1.14s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:09,  1.13s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.12s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.12s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.11s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:10<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:11<00:03,  1.10s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:12<00:02,  1.11s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.11s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.06s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.10s/it]
+Evaluated 400 examples
+[2026-07-22 17:26:09,314][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-22 17:26:09,315][evaluator][INFO] - Result for metric forget_truth_ratio:	0.5398811771455105
+[2026-07-22 17:26:09,319][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-22 17:26:09,319][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:26:09,326][metrics][INFO] - Evaluating forget_quality
+[2026-07-22 17:26:09,327][evaluator][INFO] - Result for metric forget_quality:	1.02724452984523e-14
+[2026-07-22 17:26:09,793][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.31it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.39it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.26it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:02,  4.28it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.30it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.34it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.34it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.45it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.52it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.53it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.52it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.51it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.76it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.49it/s]
+Evaluated 400 examples
+[2026-07-22 17:26:12,689][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.6085498046875
+[2026-07-22 17:26:12,996][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:04,  2.97it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:03,  2.84it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:01<00:03,  2.60it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:03,  2.67it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:03,  2.60it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:02<00:02,  2.79it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:02<00:02,  2.79it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  2.89it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:03<00:01,  2.90it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:03<00:01,  2.84it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:03<00:00,  2.75it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:04<00:00,  2.74it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:04<00:00,  2.80it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:04<00:00,  2.78it/s]
+Evaluated 400 examples
+[2026-07-22 17:26:17,675][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.46986434920230424
+[2026-07-22 17:26:18,071][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.50it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.64it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.67it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.66it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.64it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.57it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.56it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.56it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.62it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.63it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.58it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.63it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.86it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.67it/s]
+Evaluated 400 examples
+[2026-07-22 17:26:21,183][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:03,  3.11it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:03,  3.13it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:03,  3.20it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:02,  3.56it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:02,  3.71it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:01<00:02,  3.47it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:02<00:01,  3.37it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  3.51it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:02<00:01,  3.54it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:02<00:00,  3.54it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:03<00:00,  3.31it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:03<00:00,  3.29it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.62it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.46it/s]
+Evaluated 400 examples
+[2026-07-22 17:26:25,279][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.58it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.60it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.58it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.64it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.67it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.58it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.53it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.48it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.48it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.45it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.40it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.42it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.66it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.55it/s]
+Evaluated 400 examples
+[2026-07-22 17:26:28,445][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:12,  1.06s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:11,  1.06s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:10,  1.07s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:09,  1.07s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:08,  1.06s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.08s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.09s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.10s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:09<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:10<00:03,  1.10s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:11<00:02,  1.11s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.10s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.06s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.08s/it]
+Evaluated 400 examples
+[2026-07-22 17:26:42,480][metrics][INFO] - Evaluating retain_Truth_Ratio
+[2026-07-22 17:26:42,918][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.08it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.53it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  7.49it/s]
+Evaluated 100 examples
+[2026-07-22 17:26:43,763][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.34it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.45it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.06it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.47it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.50it/s]
+Evaluated 100 examples
+[2026-07-22 17:26:45,366][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+[2026-07-22 17:26:45,671][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  4.41it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  5.27it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  4.45it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  5.57it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  5.22it/s]
+Evaluated 100 examples
+[2026-07-22 17:26:46,438][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-22 17:26:46,438][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:26:46,438][metrics][INFO] - Evaluating ra_Truth_Ratio
+[2026-07-22 17:26:46,792][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.17it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.64it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  8.43it/s]
+Evaluated 117 examples
+[2026-07-22 17:26:47,580][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.77it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.51it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.09it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.56it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.58it/s]
+Evaluated 117 examples
+[2026-07-22 17:26:49,132][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+[2026-07-22 17:26:49,434][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  4.01it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  4.07it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  3.50it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  3.54it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  3.63it/s]
+Evaluated 117 examples
+[2026-07-22 17:26:50,538][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-22 17:26:50,538][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:26:50,538][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-22 17:26:50,538][metrics][INFO] - Evaluating model_utility
+[2026-07-22 17:26:50,539][evaluator][INFO] - Result for metric model_utility:	0.5846218776545004
+[2026-07-22 17:26:51,197][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:26:51,211][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 12.70it/s]
 31%|███       | 4/13 [00:00<00:00, 11.55it/s]
 46%|████▌     | 6/13 [00:00<00:00, 12.43it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 12.81it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.00it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 12.70it/s]
100%|██████████| 13/13 [00:00<00:00, 13.17it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 12.46it/s]
 31%|███       | 4/13 [00:00<00:00, 11.95it/s]
 46%|████▌     | 6/13 [00:00<00:00, 12.61it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 12.55it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 12.65it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 12.70it/s]
100%|██████████| 13/13 [00:01<00:00, 12.98it/s]
+[2026-07-22 17:26:53,203][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:26:53,212][metrics][INFO] - Evaluating privleak
+[2026-07-22 17:26:53,213][evaluator][INFO] - Result for metric privleak:	-97.5562349257871
+[2026-07-22 17:26:53,816][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:00, 11.64it/s]
Calculating ES:  31%|███       | 4/13 [00:00<00:00, 10.64it/s]
Calculating ES:  46%|████▌     | 6/13 [00:00<00:00, 11.35it/s]
Calculating ES:  62%|██████▏   | 8/13 [00:00<00:00, 11.78it/s]
Calculating ES:  77%|███████▋  | 10/13 [00:00<00:00, 11.93it/s]
Calculating ES:  92%|█████████▏| 12/13 [00:01<00:00, 11.69it/s]
Calculating ES: 100%|██████████| 13/13 [00:01<00:00, 12.09it/s]
+Evaluated 400 examples
+[2026-07-22 17:26:54,891][evaluator][INFO] - Result for metric extraction_strength:	0.20012088248929616
diff --git a/logs/eval_NPO_lr1e-05_beta0.5_alpha2_epoch10.log b/logs/eval_NPO_lr1e-05_beta0.5_alpha2_epoch10.log
new file mode 100644
index 0000000000000000000000000000000000000000..1b4703b78e54ce044d98f744ff0ff56b88fe0f71
--- /dev/null
+++ b/logs/eval_NPO_lr1e-05_beta0.5_alpha2_epoch10.log
@@ -0,0 +1,78 @@
+/opt/conda/envs/copygen/lib/python3.12/site-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.6.3) or chardet (6.0.0.post1)/charset_normalizer (3.4.4) doesn't match a supported version!
+  warnings.warn(
+[2026-07-22 17:23:40,146][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/NPO_lr1e-05_beta0.5_alpha2_epoch10
+[2026-07-22 17:23:40,147][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-22 17:23:40,147][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/NPO_lr1e-05_beta0.5_alpha2_epoch10/TOFU_EVAL.json
+[2026-07-22 17:23:40,147][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/NPO_lr1e-05_beta0.5_alpha2_epoch10/TOFU_SUMMARY.json
+[2026-07-22 17:23:40,955][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:05,  2.12it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:03,  3.01it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  3.45it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.74it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:02,  3.97it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.10it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.21it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:02<00:01,  4.34it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.39it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.43it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.42it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.40it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.60it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.11it/s]
+Evaluated 400 examples
+[2026-07-22 17:23:44,529][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:13,  1.12s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:12,  1.12s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:11,  1.15s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:10,  1.15s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:09,  1.15s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.14s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.13s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:09<00:05,  1.11s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:10<00:04,  1.11s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:11<00:03,  1.11s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:12<00:02,  1.12s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.12s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.07s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.11s/it]
+Evaluated 400 examples
+[2026-07-22 17:23:58,985][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-22 17:23:58,986][evaluator][INFO] - Result for metric forget_truth_ratio:	0.5154405212249733
+[2026-07-22 17:23:58,990][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-22 17:23:58,990][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:23:58,997][metrics][INFO] - Evaluating forget_quality
+[2026-07-22 17:23:58,998][evaluator][INFO] - Result for metric forget_quality:	1.730834713171486e-15
+[2026-07-22 17:24:04,594][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.47it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.44it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.28it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:02,  4.32it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.35it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.39it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.38it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.47it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.47it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.44it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.40it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.38it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.64it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.46it/s]
+Evaluated 400 examples
+[2026-07-22 17:24:07,511][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.57855224609375
+[2026-07-22 17:24:07,871][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:03,  3.05it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:03,  3.01it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:01<00:03,  2.65it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:03,  2.42it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:03,  2.56it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:02<00:02,  2.86it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:02<00:02,  2.99it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  3.20it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:03<00:01,  3.05it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:03<00:01,  2.89it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:03<00:00,  2.99it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:04<00:00,  2.76it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:04<00:00,  2.87it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:04<00:00,  2.86it/s]
+Evaluated 400 examples
+[2026-07-22 17:24:12,420][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.4641889517967577
+[2026-07-22 17:24:12,766][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.62it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.67it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.68it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.68it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.69it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.64it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.64it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.59it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.64it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.58it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.50it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.53it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.75it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.65it/s]
+Evaluated 400 examples
+[2026-07-22 17:24:15,861][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:04,  2.82it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:03,  3.14it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:02,  3.37it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:02,  3.64it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:02,  3.68it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:01<00:01,  3.65it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:02<00:01,  3.37it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  3.27it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:02<00:01,  3.37it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:02<00:00,  3.43it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:03<00:00,  3.12it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:03<00:00,  3.19it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.46it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.38it/s]
+Evaluated 400 examples
+[2026-07-22 17:24:20,016][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.59it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.59it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.55it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.61it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.64it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.56it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.52it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.47it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.47it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.49it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.46it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.49it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.70it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.57it/s]
+Evaluated 400 examples
+[2026-07-22 17:24:23,232][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:12,  1.07s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:11,  1.06s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:10,  1.08s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:09,  1.07s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:08,  1.06s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.08s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.09s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.10s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:09<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:10<00:03,  1.10s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:11<00:02,  1.10s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.10s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.05s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.08s/it]
+Evaluated 400 examples
+[2026-07-22 17:24:37,248][metrics][INFO] - Evaluating retain_Truth_Ratio
+[2026-07-22 17:24:37,701][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.13it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.54it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  7.49it/s]
+Evaluated 100 examples
+[2026-07-22 17:24:38,623][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.37it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.47it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.05it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.37it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.44it/s]
+Evaluated 100 examples
+[2026-07-22 17:24:40,265][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+[2026-07-22 17:24:40,607][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  4.23it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  4.99it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  5.12it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  6.29it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  5.68it/s]
+Evaluated 100 examples
+[2026-07-22 17:24:41,312][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-22 17:24:41,312][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:24:41,312][metrics][INFO] - Evaluating ra_Truth_Ratio
+[2026-07-22 17:24:41,660][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.18it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.64it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  8.44it/s]
+Evaluated 117 examples
+[2026-07-22 17:24:42,438][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.85it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.51it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.11it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.56it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.59it/s]
+Evaluated 117 examples
+[2026-07-22 17:24:43,986][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+[2026-07-22 17:24:44,312][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  3.89it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  4.04it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  4.32it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  4.29it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  4.22it/s]
+Evaluated 117 examples
+[2026-07-22 17:24:45,260][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-22 17:24:45,260][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:24:45,260][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-22 17:24:45,260][metrics][INFO] - Evaluating model_utility
+[2026-07-22 17:24:45,261][evaluator][INFO] - Result for metric model_utility:	0.5938613153598361
+[2026-07-22 17:24:45,991][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:24:45,999][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.67it/s]
 31%|███       | 4/13 [00:00<00:00, 12.40it/s]
 46%|████▌     | 6/13 [00:00<00:00, 12.83it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 12.97it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.46it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.26it/s]
100%|██████████| 13/13 [00:00<00:00, 13.70it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 12.84it/s]
 31%|███       | 4/13 [00:00<00:00, 12.51it/s]
 46%|████▌     | 6/13 [00:00<00:00, 12.96it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 12.96it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.00it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 12.77it/s]
100%|██████████| 13/13 [00:00<00:00, 13.14it/s]
+[2026-07-22 17:24:47,940][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:24:47,948][metrics][INFO] - Evaluating privleak
+[2026-07-22 17:24:47,948][evaluator][INFO] - Result for metric privleak:	-97.58862949159341
+[2026-07-22 17:24:48,290][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:00, 11.74it/s]
Calculating ES:  31%|███       | 4/13 [00:00<00:00, 10.82it/s]
Calculating ES:  46%|████▌     | 6/13 [00:00<00:00, 11.54it/s]
Calculating ES:  62%|██████▏   | 8/13 [00:00<00:00, 11.98it/s]
Calculating ES:  77%|███████▋  | 10/13 [00:00<00:00, 12.02it/s]
Calculating ES:  92%|█████████▏| 12/13 [00:01<00:00, 11.81it/s]
Calculating ES: 100%|██████████| 13/13 [00:01<00:00, 12.19it/s]
+Evaluated 400 examples
+[2026-07-22 17:24:49,357][evaluator][INFO] - Result for metric extraction_strength:	0.19484571565832418
diff --git a/logs/eval_NPO_lr1e-05_beta0.5_alpha2_epoch5.log b/logs/eval_NPO_lr1e-05_beta0.5_alpha2_epoch5.log
new file mode 100644
index 0000000000000000000000000000000000000000..686193d302ddfca90b39cf0e3ba923928a41b706
--- /dev/null
+++ b/logs/eval_NPO_lr1e-05_beta0.5_alpha2_epoch5.log
@@ -0,0 +1,78 @@
+/opt/conda/envs/copygen/lib/python3.12/site-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.6.3) or chardet (6.0.0.post1)/charset_normalizer (3.4.4) doesn't match a supported version!
+  warnings.warn(
+[2026-07-22 17:31:32,714][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/NPO_lr1e-05_beta0.5_alpha2_epoch5
+[2026-07-22 17:31:32,715][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-22 17:31:32,715][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/NPO_lr1e-05_beta0.5_alpha2_epoch5/TOFU_EVAL.json
+[2026-07-22 17:31:32,715][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/NPO_lr1e-05_beta0.5_alpha2_epoch5/TOFU_SUMMARY.json
+[2026-07-22 17:31:33,506][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:05,  2.26it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:03,  3.17it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  3.56it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.80it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.01it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.11it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.21it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:02<00:01,  4.34it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.34it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.35it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.33it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.30it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.52it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.11it/s]
+Evaluated 400 examples
+[2026-07-22 17:31:37,150][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:13,  1.12s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:12,  1.11s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:11,  1.14s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:10,  1.15s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:09,  1.13s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.13s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.13s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.11s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:10<00:04,  1.11s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:11<00:03,  1.10s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:12<00:02,  1.11s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.12s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.06s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.11s/it]
+Evaluated 400 examples
+[2026-07-22 17:31:51,541][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-22 17:31:51,541][evaluator][INFO] - Result for metric forget_truth_ratio:	0.5228238653855537
+[2026-07-22 17:31:51,545][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-22 17:31:51,546][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:31:51,554][metrics][INFO] - Evaluating forget_quality
+[2026-07-22 17:31:51,554][evaluator][INFO] - Result for metric forget_quality:	9.456104707095184e-16
+[2026-07-22 17:31:52,046][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.43it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.45it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.26it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:02,  4.26it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.29it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.34it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.35it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.45it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.51it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.52it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.50it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.49it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.73it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.48it/s]
+Evaluated 400 examples
+[2026-07-22 17:31:54,946][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.6081005859375
+[2026-07-22 17:31:55,259][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:03,  3.02it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:03,  2.97it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:01<00:03,  2.63it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:03,  2.63it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:02,  2.87it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:02<00:02,  3.09it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:02<00:01,  3.23it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  3.20it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:03<00:01,  3.02it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:03<00:01,  2.87it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:03<00:00,  2.88it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:04<00:00,  2.80it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:04<00:00,  2.97it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:04<00:00,  2.94it/s]
+Evaluated 400 examples
+[2026-07-22 17:31:59,685][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.4660188810798068
+[2026-07-22 17:32:00,009][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.57it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.66it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.68it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.71it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.69it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.58it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.54it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.51it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.56it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.58it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.55it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.60it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.82it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.65it/s]
+Evaluated 400 examples
+[2026-07-22 17:32:03,121][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:03,  3.05it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:03,  3.53it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:02,  3.55it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:02,  3.78it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:02,  3.48it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:01<00:02,  3.48it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:02<00:01,  3.25it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  3.32it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:02<00:01,  3.42it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:02<00:00,  3.16it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:03<00:00,  3.14it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:03<00:00,  3.21it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.49it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.38it/s]
+Evaluated 400 examples
+[2026-07-22 17:32:07,278][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.59it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.60it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.57it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.61it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.61it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.54it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.51it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.47it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.47it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.47it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.41it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.42it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.65it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.54it/s]
+Evaluated 400 examples
+[2026-07-22 17:32:10,439][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:12,  1.06s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:11,  1.06s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:10,  1.07s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:09,  1.07s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:08,  1.06s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.08s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.09s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.10s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:09<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:10<00:03,  1.09s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:11<00:02,  1.10s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.10s/it]
Calculating loss: 100%|██████████| 13/13 [00:13<00:00,  1.05s/it]
Calculating loss: 100%|██████████| 13/13 [00:13<00:00,  1.08s/it]
+Evaluated 400 examples
+[2026-07-22 17:32:24,430][metrics][INFO] - Evaluating retain_Truth_Ratio
+[2026-07-22 17:32:24,843][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  7.78it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.30it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  7.21it/s]
+Evaluated 100 examples
+[2026-07-22 17:32:25,698][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.22it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.43it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.06it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.47it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.50it/s]
+Evaluated 100 examples
+[2026-07-22 17:32:27,300][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+[2026-07-22 17:32:27,600][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  4.79it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  5.68it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  5.29it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  6.42it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  5.96it/s]
+Evaluated 100 examples
+[2026-07-22 17:32:28,271][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-22 17:32:28,272][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:32:28,272][metrics][INFO] - Evaluating ra_Truth_Ratio
+[2026-07-22 17:32:28,581][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.18it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.62it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  8.41it/s]
+Evaluated 117 examples
+[2026-07-22 17:32:29,405][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.90it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.51it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.10it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.57it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.58it/s]
+Evaluated 117 examples
+[2026-07-22 17:32:30,954][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+[2026-07-22 17:32:31,260][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  4.81it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  4.45it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  4.05it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  4.35it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  4.34it/s]
+Evaluated 117 examples
+[2026-07-22 17:32:32,181][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-22 17:32:32,181][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:32:32,181][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-22 17:32:32,181][metrics][INFO] - Evaluating model_utility
+[2026-07-22 17:32:32,182][evaluator][INFO] - Result for metric model_utility:	0.5892532915319522
+[2026-07-22 17:32:32,839][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:32:32,847][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.36it/s]
 31%|███       | 4/13 [00:00<00:00, 12.34it/s]
 46%|████▌     | 6/13 [00:00<00:00, 13.09it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.02it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.59it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.25it/s]
100%|██████████| 13/13 [00:00<00:00, 13.75it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.54it/s]
 31%|███       | 4/13 [00:00<00:00, 13.33it/s]
 46%|████▌     | 6/13 [00:00<00:00, 13.70it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.42it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.44it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.53it/s]
100%|██████████| 13/13 [00:00<00:00, 13.93it/s]
+[2026-07-22 17:32:34,728][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:32:34,736][metrics][INFO] - Evaluating privleak
+[2026-07-22 17:32:34,736][evaluator][INFO] - Result for metric privleak:	-97.83057640495953
+[2026-07-22 17:32:35,098][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:00, 11.80it/s]
Calculating ES:  31%|███       | 4/13 [00:00<00:00, 10.83it/s]
Calculating ES:  46%|████▌     | 6/13 [00:00<00:00, 11.49it/s]
Calculating ES:  62%|██████▏   | 8/13 [00:00<00:00, 11.93it/s]
Calculating ES:  77%|███████▋  | 10/13 [00:00<00:00, 12.11it/s]
Calculating ES:  92%|█████████▏| 12/13 [00:01<00:00, 11.74it/s]
Calculating ES: 100%|██████████| 13/13 [00:01<00:00, 12.20it/s]
+Evaluated 400 examples
+[2026-07-22 17:32:36,164][evaluator][INFO] - Result for metric extraction_strength:	0.20613616336171042
diff --git a/logs/eval_NPO_lr1e-05_beta0.5_alpha5_epoch10.log b/logs/eval_NPO_lr1e-05_beta0.5_alpha5_epoch10.log
new file mode 100644
index 0000000000000000000000000000000000000000..3d779b02aa69931fe8d3befa7a8165f23b349a62
--- /dev/null
+++ b/logs/eval_NPO_lr1e-05_beta0.5_alpha5_epoch10.log
@@ -0,0 +1,78 @@
+/opt/conda/envs/copygen/lib/python3.12/site-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.6.3) or chardet (6.0.0.post1)/charset_normalizer (3.4.4) doesn't match a supported version!
+  warnings.warn(
+[2026-07-22 17:27:20,446][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/NPO_lr1e-05_beta0.5_alpha5_epoch10
+[2026-07-22 17:27:20,447][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-22 17:27:20,447][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/NPO_lr1e-05_beta0.5_alpha5_epoch10/TOFU_EVAL.json
+[2026-07-22 17:27:20,447][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/NPO_lr1e-05_beta0.5_alpha5_epoch10/TOFU_SUMMARY.json
+[2026-07-22 17:27:21,207][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:04,  2.46it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:03,  3.33it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  3.68it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.91it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.07it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.16it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.25it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:02<00:01,  4.36it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.40it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.44it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.42it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.36it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.57it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.19it/s]
+Evaluated 400 examples
+[2026-07-22 17:27:24,793][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:13,  1.12s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:12,  1.13s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:11,  1.15s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:10,  1.14s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:09,  1.13s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.13s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.12s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.10s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:10<00:04,  1.11s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:11<00:03,  1.10s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:12<00:02,  1.11s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.11s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.06s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.10s/it]
+Evaluated 400 examples
+[2026-07-22 17:27:39,141][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-22 17:27:39,142][evaluator][INFO] - Result for metric forget_truth_ratio:	0.5032124607036412
+[2026-07-22 17:27:39,146][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-22 17:27:39,146][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:27:39,153][metrics][INFO] - Evaluating forget_quality
+[2026-07-22 17:27:39,154][evaluator][INFO] - Result for metric forget_quality:	2.776495661774257e-16
+[2026-07-22 17:27:39,577][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.42it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.42it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.29it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:02,  4.33it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.37it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.42it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.42it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.51it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.55it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.53it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.51it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.46it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.67it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.50it/s]
+Evaluated 400 examples
+[2026-07-22 17:27:42,466][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.6050927734375
+[2026-07-22 17:27:42,774][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:04,  2.75it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:04,  2.50it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:01<00:04,  2.32it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:04,  2.15it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:02<00:03,  2.30it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:02<00:02,  2.66it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:02<00:02,  2.63it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:03<00:01,  2.81it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:03<00:01,  2.94it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:03<00:01,  2.88it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:04<00:00,  2.93it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:04<00:00,  2.83it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:04<00:00,  2.95it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:04<00:00,  2.71it/s]
+Evaluated 400 examples
+[2026-07-22 17:27:47,569][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.4963806389208682
+[2026-07-22 17:27:47,929][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.63it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.67it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.69it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.72it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.71it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.64it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.64it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.61it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.65it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.64it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.58it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.60it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.78it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.68it/s]
+Evaluated 400 examples
+[2026-07-22 17:27:51,073][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:03,  3.22it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:03,  3.31it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:03,  3.33it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:02,  3.61it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:02,  3.69it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:01<00:01,  3.63it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:02<00:01,  3.37it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  3.34it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:02<00:01,  3.29it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:02<00:00,  3.29it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:03<00:00,  3.07it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:03<00:00,  3.18it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.37it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.35it/s]
+Evaluated 400 examples
+[2026-07-22 17:27:55,255][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.51it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.57it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.56it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.62it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.66it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.58it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.55it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.49it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.48it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.49it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.46it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.49it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.73it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.58it/s]
+Evaluated 400 examples
+[2026-07-22 17:27:58,398][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:13,  1.09s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:11,  1.07s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:10,  1.08s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:09,  1.07s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:08,  1.06s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.08s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.09s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.10s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:09<00:04,  1.11s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:10<00:03,  1.10s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:12<00:02,  1.11s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.10s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.05s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.08s/it]
+Evaluated 400 examples
+[2026-07-22 17:28:12,427][metrics][INFO] - Evaluating retain_Truth_Ratio
+[2026-07-22 17:28:12,823][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.11it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.53it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  7.50it/s]
+Evaluated 100 examples
+[2026-07-22 17:28:13,674][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.30it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.48it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.08it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.46it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.50it/s]
+Evaluated 100 examples
+[2026-07-22 17:28:15,274][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+[2026-07-22 17:28:15,578][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  3.52it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  4.12it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  4.30it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  4.91it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  4.55it/s]
+Evaluated 100 examples
+[2026-07-22 17:28:16,457][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-22 17:28:16,457][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:28:16,458][metrics][INFO] - Evaluating ra_Truth_Ratio
+[2026-07-22 17:28:16,821][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.09it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.59it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  8.34it/s]
+Evaluated 117 examples
+[2026-07-22 17:28:17,627][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.72it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.50it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.10it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.57it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.58it/s]
+Evaluated 117 examples
+[2026-07-22 17:28:19,177][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+[2026-07-22 17:28:19,531][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  3.67it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  3.78it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  4.06it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  3.96it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  3.93it/s]
+Evaluated 117 examples
+[2026-07-22 17:28:20,550][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-22 17:28:20,550][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:28:20,550][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-22 17:28:20,550][metrics][INFO] - Evaluating model_utility
+[2026-07-22 17:28:20,550][evaluator][INFO] - Result for metric model_utility:	0.5956605056878908
+[2026-07-22 17:28:22,076][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:28:22,084][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.23it/s]
 31%|███       | 4/13 [00:00<00:00, 12.00it/s]
 46%|████▌     | 6/13 [00:00<00:00, 12.46it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 12.78it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.08it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 12.76it/s]
100%|██████████| 13/13 [00:00<00:00, 13.26it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 12.62it/s]
 31%|███       | 4/13 [00:00<00:00, 12.41it/s]
 46%|████▌     | 6/13 [00:00<00:00, 12.81it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 12.88it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 12.82it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 12.94it/s]
100%|██████████| 13/13 [00:00<00:00, 13.27it/s]
+[2026-07-22 17:28:24,047][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:28:24,055][metrics][INFO] - Evaluating privleak
+[2026-07-22 17:28:24,055][evaluator][INFO] - Result for metric privleak:	-97.86904495185455
+[2026-07-22 17:28:24,379][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:00, 11.23it/s]
Calculating ES:  31%|███       | 4/13 [00:00<00:00, 10.30it/s]
Calculating ES:  46%|████▌     | 6/13 [00:00<00:00, 10.96it/s]
Calculating ES:  62%|██████▏   | 8/13 [00:00<00:00, 11.34it/s]
Calculating ES:  77%|███████▋  | 10/13 [00:00<00:00, 11.62it/s]
Calculating ES:  92%|█████████▏| 12/13 [00:01<00:00, 11.50it/s]
Calculating ES: 100%|██████████| 13/13 [00:01<00:00, 11.79it/s]
+Evaluated 400 examples
+[2026-07-22 17:28:25,483][evaluator][INFO] - Result for metric extraction_strength:	0.21604010964407763
diff --git a/logs/eval_NPO_lr1e-05_beta0.5_alpha5_epoch5.log b/logs/eval_NPO_lr1e-05_beta0.5_alpha5_epoch5.log
new file mode 100644
index 0000000000000000000000000000000000000000..f293edaad95faa470c3b187f1ea8cac175aec963
--- /dev/null
+++ b/logs/eval_NPO_lr1e-05_beta0.5_alpha5_epoch5.log
@@ -0,0 +1,78 @@
+/opt/conda/envs/copygen/lib/python3.12/site-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.6.3) or chardet (6.0.0.post1)/charset_normalizer (3.4.4) doesn't match a supported version!
+  warnings.warn(
+[2026-07-22 17:25:17,078][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/NPO_lr1e-05_beta0.5_alpha5_epoch5
+[2026-07-22 17:25:17,079][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-22 17:25:17,079][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/NPO_lr1e-05_beta0.5_alpha5_epoch5/TOFU_EVAL.json
+[2026-07-22 17:25:17,079][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/NPO_lr1e-05_beta0.5_alpha5_epoch5/TOFU_SUMMARY.json
+[2026-07-22 17:25:17,876][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:05,  2.19it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:03,  3.10it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  3.51it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.76it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:02,  3.97it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.08it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.19it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:02<00:01,  4.32it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.37it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.42it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.39it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.37it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.53it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.10it/s]
+Evaluated 400 examples
+[2026-07-22 17:25:21,595][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:13,  1.13s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:12,  1.12s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:11,  1.15s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:10,  1.15s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:09,  1.13s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.13s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.13s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:09<00:05,  1.12s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:10<00:04,  1.11s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:11<00:03,  1.11s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:12<00:02,  1.11s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.12s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.06s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.11s/it]
+Evaluated 400 examples
+[2026-07-22 17:25:36,011][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-22 17:25:36,012][evaluator][INFO] - Result for metric forget_truth_ratio:	0.5021398655432685
+[2026-07-22 17:25:36,017][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-22 17:25:36,017][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:25:36,025][metrics][INFO] - Evaluating forget_quality
+[2026-07-22 17:25:36,027][evaluator][INFO] - Result for metric forget_quality:	7.975373070421785e-17
+[2026-07-22 17:25:36,437][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.40it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.42it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.27it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:02,  4.28it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.34it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.37it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.38it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.42it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.45it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.43it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.42it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.43it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.69it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.46it/s]
+Evaluated 400 examples
+[2026-07-22 17:25:39,352][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.611793212890625
+[2026-07-22 17:25:39,780][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:03,  3.01it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:03,  2.97it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:01<00:03,  2.55it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:03,  2.41it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:03,  2.50it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:02<00:02,  2.61it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:02<00:02,  2.61it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  2.78it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:03<00:01,  2.91it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:03<00:01,  2.71it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:04<00:00,  2.72it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:04<00:00,  2.74it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:04<00:00,  2.95it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:04<00:00,  2.75it/s]
+Evaluated 400 examples
+[2026-07-22 17:25:44,515][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.48653544535250126
+[2026-07-22 17:25:44,911][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.65it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.71it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.71it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.73it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.73it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.66it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.58it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.53it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.54it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.54it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.51it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.56it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.80it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.65it/s]
+Evaluated 400 examples
+[2026-07-22 17:25:48,123][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:03,  3.31it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:03,  3.27it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:02,  3.36it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:02,  3.63it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:02,  3.73it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:01<00:01,  3.67it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:02<00:01,  3.38it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  3.28it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:02<00:01,  3.30it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:02<00:00,  3.19it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:03<00:00,  3.07it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:03<00:00,  3.15it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.40it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.36it/s]
+Evaluated 400 examples
+[2026-07-22 17:25:52,841][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.50it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.56it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.53it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.60it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.63it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.55it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.53it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.42it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.40it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.41it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.39it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.44it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.70it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.54it/s]
+Evaluated 400 examples
+[2026-07-22 17:25:56,021][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:12,  1.06s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:11,  1.06s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:10,  1.07s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:09,  1.07s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:08,  1.06s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.07s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.09s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.10s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:09<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:10<00:03,  1.09s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:11<00:02,  1.10s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.10s/it]
Calculating loss: 100%|██████████| 13/13 [00:13<00:00,  1.05s/it]
Calculating loss: 100%|██████████| 13/13 [00:13<00:00,  1.08s/it]
+Evaluated 400 examples
+[2026-07-22 17:26:10,014][metrics][INFO] - Evaluating retain_Truth_Ratio
+[2026-07-22 17:26:10,424][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.10it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.45it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  6.53it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  6.71it/s]
+Evaluated 100 examples
+[2026-07-22 17:26:11,332][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.03it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.47it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.09it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.48it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.51it/s]
+Evaluated 100 examples
+[2026-07-22 17:26:12,923][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+[2026-07-22 17:26:13,220][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  4.22it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  5.33it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  5.02it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  5.67it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  5.37it/s]
+Evaluated 100 examples
+[2026-07-22 17:26:13,965][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-22 17:26:13,965][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:26:13,965][metrics][INFO] - Evaluating ra_Truth_Ratio
+[2026-07-22 17:26:14,262][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.17it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.63it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  8.40it/s]
+Evaluated 117 examples
+[2026-07-22 17:26:15,037][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.50it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.46it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.08it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.55it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.56it/s]
+Evaluated 117 examples
+[2026-07-22 17:26:16,600][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+[2026-07-22 17:26:16,911][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  3.78it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  3.83it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  3.18it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  3.26it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  3.34it/s]
+Evaluated 117 examples
+[2026-07-22 17:26:18,108][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-22 17:26:18,108][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:26:18,108][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-22 17:26:18,108][metrics][INFO] - Evaluating model_utility
+[2026-07-22 17:26:18,108][evaluator][INFO] - Result for metric model_utility:	0.5925075559313268
+[2026-07-22 17:26:18,753][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:26:18,761][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.33it/s]
 31%|███       | 4/13 [00:00<00:00, 12.05it/s]
 46%|████▌     | 6/13 [00:00<00:00, 12.65it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.00it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.38it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.23it/s]
100%|██████████| 13/13 [00:00<00:00, 13.63it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.28it/s]
 31%|███       | 4/13 [00:00<00:00, 13.10it/s]
 46%|████▌     | 6/13 [00:00<00:00, 12.51it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 12.16it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 12.42it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 12.87it/s]
100%|██████████| 13/13 [00:00<00:00, 13.18it/s]
+[2026-07-22 17:26:20,703][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:26:20,710][metrics][INFO] - Evaluating privleak
+[2026-07-22 17:26:20,710][evaluator][INFO] - Result for metric privleak:	-97.9095381591125
+[2026-07-22 17:26:21,114][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:00, 11.87it/s]
Calculating ES:  31%|███       | 4/13 [00:00<00:00, 10.93it/s]
Calculating ES:  46%|████▌     | 6/13 [00:00<00:00, 11.61it/s]
Calculating ES:  62%|██████▏   | 8/13 [00:00<00:00, 12.05it/s]
Calculating ES:  77%|███████▋  | 10/13 [00:00<00:00, 12.22it/s]
Calculating ES:  92%|█████████▏| 12/13 [00:01<00:00, 11.96it/s]
Calculating ES: 100%|██████████| 13/13 [00:01<00:00, 12.37it/s]
+Evaluated 400 examples
+[2026-07-22 17:26:22,165][evaluator][INFO] - Result for metric extraction_strength:	0.22206953870624216
diff --git a/logs/eval_NPO_lr2e-05_beta0.05_alpha1_epoch10.log b/logs/eval_NPO_lr2e-05_beta0.05_alpha1_epoch10.log
new file mode 100644
index 0000000000000000000000000000000000000000..7f8c2f209602844fda562d7e9fba61efd0d699f0
--- /dev/null
+++ b/logs/eval_NPO_lr2e-05_beta0.05_alpha1_epoch10.log
@@ -0,0 +1,78 @@
+/opt/conda/envs/copygen/lib/python3.12/site-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.6.3) or chardet (6.0.0.post1)/charset_normalizer (3.4.4) doesn't match a supported version!
+  warnings.warn(
+[2026-07-20 13:30:32,508][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/NPO_lr2e-05_beta0.05_alpha1_epoch10
+[2026-07-20 13:30:32,509][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-20 13:30:32,509][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/NPO_lr2e-05_beta0.05_alpha1_epoch10/TOFU_EVAL.json
+[2026-07-20 13:30:32,509][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/NPO_lr2e-05_beta0.05_alpha1_epoch10/TOFU_SUMMARY.json
+[2026-07-20 13:30:33,384][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:08,  1.48it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:04,  2.48it/s]
Calculating loss:  23%|██▎       | 3/13 [00:01<00:03,  3.12it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.59it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:02,  3.93it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.15it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.34it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:02<00:01,  4.52it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.53it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.59it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.62it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:03<00:00,  4.61it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.78it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.04it/s]
+Evaluated 400 examples
+[2026-07-20 13:30:36,970][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:12,  1.03s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:11,  1.02s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:10,  1.05s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:09,  1.05s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:08,  1.04s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.03s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.03s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.01s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:09<00:04,  1.01s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:10<00:03,  1.02s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:11<00:02,  1.02s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:12<00:01,  1.02s/it]
Calculating loss: 100%|██████████| 13/13 [00:13<00:00,  1.03it/s]
Calculating loss: 100%|██████████| 13/13 [00:13<00:00,  1.01s/it]
+Evaluated 400 examples
+[2026-07-20 13:30:50,155][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-20 13:30:50,156][evaluator][INFO] - Result for metric forget_truth_ratio:	0.5682724549448078
+[2026-07-20 13:30:50,165][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-20 13:30:50,166][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-20 13:30:50,182][metrics][INFO] - Evaluating forget_quality
+[2026-07-20 13:30:50,184][evaluator][INFO] - Result for metric forget_quality:	0.0022938851536275073
+[2026-07-20 13:30:50,624][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.76it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.79it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.68it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.71it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.73it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.77it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.77it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.85it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.89it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.86it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.85it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.80it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  5.06it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.86it/s]
+Evaluated 400 examples
+[2026-07-20 13:30:53,299][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.015265593096555676
+[2026-07-20 13:30:53,604][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:08,  1.45it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:01<00:06,  1.60it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:02<00:06,  1.43it/s]
Calculating text similarity:  31%|███       | 4/13 [00:02<00:06,  1.44it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:03<00:04,  1.62it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:03<00:04,  1.59it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:04<00:03,  1.53it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:05<00:03,  1.57it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:05<00:02,  1.54it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:06<00:01,  1.51it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:07<00:01,  1.50it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:07<00:00,  1.54it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:08<00:00,  1.66it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:08<00:00,  1.56it/s]
+Evaluated 400 examples
+[2026-07-20 13:31:01,944][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.24811344076238828
+[2026-07-20 13:31:02,313][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.89it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.97it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.99it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  5.02it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  5.01it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.94it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.93it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.90it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.93it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.93it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.89it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.93it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  5.17it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  5.00it/s]
+Evaluated 400 examples
+[2026-07-20 13:31:05,220][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:07,  1.60it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:01<00:06,  1.70it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:01<00:06,  1.57it/s]
Calculating text similarity:  31%|███       | 4/13 [00:02<00:05,  1.51it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:03<00:04,  1.62it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:03<00:04,  1.61it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:04<00:03,  1.58it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:05<00:03,  1.41it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:05<00:02,  1.44it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:06<00:02,  1.44it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:07<00:01,  1.43it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:08<00:00,  1.44it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:08<00:00,  1.53it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:08<00:00,  1.51it/s]
+Evaluated 400 examples
+[2026-07-20 13:31:14,198][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.37it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.69it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.77it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.88it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.92it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.87it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.83it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.80it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.80it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.82it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.79it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.82it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  5.09it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.87it/s]
+Evaluated 400 examples
+[2026-07-20 13:31:17,314][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:11,  1.02it/s]
Calculating loss:  15%|█▌        | 2/13 [00:01<00:10,  1.02it/s]
Calculating loss:  23%|██▎       | 3/13 [00:02<00:09,  1.01it/s]
Calculating loss:  31%|███       | 4/13 [00:03<00:08,  1.02it/s]
Calculating loss:  38%|███▊      | 5/13 [00:04<00:07,  1.02it/s]
Calculating loss:  46%|████▌     | 6/13 [00:05<00:06,  1.01it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:06<00:05,  1.00it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:07<00:05,  1.00s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:08<00:04,  1.00s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:09<00:03,  1.01s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:10<00:02,  1.01s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:11<00:01,  1.01s/it]
Calculating loss: 100%|██████████| 13/13 [00:12<00:00,  1.03it/s]
Calculating loss: 100%|██████████| 13/13 [00:12<00:00,  1.01it/s]
+Evaluated 400 examples
+[2026-07-20 13:31:30,181][metrics][INFO] - Evaluating retain_Truth_Ratio
+[2026-07-20 13:31:30,645][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.33it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.78it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  7.92it/s]
+Evaluated 100 examples
+[2026-07-20 13:31:31,407][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.24it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.62it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.21it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.62it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.66it/s]
+Evaluated 100 examples
+[2026-07-20 13:31:32,914][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+[2026-07-20 13:31:33,212][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:01,  2.20it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  2.23it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:01<00:00,  2.33it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  2.27it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  2.27it/s]
+Evaluated 100 examples
+[2026-07-20 13:31:34,977][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-20 13:31:34,977][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-20 13:31:34,977][metrics][INFO] - Evaluating ra_Truth_Ratio
+[2026-07-20 13:31:35,225][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.34it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.88it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  8.61it/s]
+Evaluated 117 examples
+[2026-07-20 13:31:35,953][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.48it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.62it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.20it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.69it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.70it/s]
+Evaluated 117 examples
+[2026-07-20 13:31:37,437][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+[2026-07-20 13:31:37,748][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:01,  2.43it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  2.48it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:01<00:00,  2.02it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  2.20it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  2.21it/s]
+Evaluated 117 examples
+[2026-07-20 13:31:39,556][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-20 13:31:39,556][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-20 13:31:39,557][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-20 13:31:39,557][metrics][INFO] - Evaluating model_utility
+[2026-07-20 13:31:39,558][evaluator][INFO] - Result for metric model_utility:	0.561656642269312
+[2026-07-20 13:31:40,250][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-20 13:31:40,265][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 12.74it/s]
 31%|███       | 4/13 [00:00<00:00, 11.31it/s]
 46%|████▌     | 6/13 [00:00<00:00, 11.69it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 11.84it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 10.47it/s]
 92%|█████████▏| 12/13 [00:01<00:00, 11.05it/s]
100%|██████████| 13/13 [00:01<00:00, 11.73it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:01,  9.77it/s]
 23%|██▎       | 3/13 [00:00<00:01,  8.48it/s]
 38%|███▊      | 5/13 [00:00<00:00, 10.19it/s]
 54%|█████▍    | 7/13 [00:00<00:00, 11.20it/s]
 69%|██████▉   | 9/13 [00:00<00:00, 11.22it/s]
 85%|████████▍ | 11/13 [00:00<00:00, 11.67it/s]
100%|██████████| 13/13 [00:01<00:00, 12.69it/s]
100%|██████████| 13/13 [00:01<00:00, 11.50it/s]
+[2026-07-20 13:31:42,508][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-20 13:31:42,522][metrics][INFO] - Evaluating privleak
+[2026-07-20 13:31:42,523][evaluator][INFO] - Result for metric privleak:	59.0046769558819
+[2026-07-20 13:31:42,921][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:01, 10.30it/s]
Calculating ES:  31%|███       | 4/13 [00:00<00:00,  9.33it/s]
Calculating ES:  38%|███▊      | 5/13 [00:00<00:00,  9.49it/s]
Calculating ES:  54%|█████▍    | 7/13 [00:00<00:00,  9.80it/s]
Calculating ES:  69%|██████▉   | 9/13 [00:00<00:00, 10.20it/s]
Calculating ES:  85%|████████▍ | 11/13 [00:01<00:00, 10.05it/s]
Calculating ES: 100%|██████████| 13/13 [00:01<00:00, 10.83it/s]
Calculating ES: 100%|██████████| 13/13 [00:01<00:00, 10.28it/s]
+Evaluated 400 examples
+[2026-07-20 13:31:44,187][evaluator][INFO] - Result for metric extraction_strength:	0.08647805985922113
diff --git a/logs/eval_NPO_lr2e-05_beta0.05_alpha1_epoch5.log b/logs/eval_NPO_lr2e-05_beta0.05_alpha1_epoch5.log
new file mode 100644
index 0000000000000000000000000000000000000000..b3dd36cb3a4acc53f3c15b615315675f0e9cf913
--- /dev/null
+++ b/logs/eval_NPO_lr2e-05_beta0.05_alpha1_epoch5.log
@@ -0,0 +1,78 @@
+/opt/conda/envs/copygen/lib/python3.12/site-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.6.3) or chardet (6.0.0.post1)/charset_normalizer (3.4.4) doesn't match a supported version!
+  warnings.warn(
+[2026-07-22 17:33:05,184][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/NPO_lr2e-05_beta0.05_alpha1_epoch5
+[2026-07-22 17:33:05,185][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-22 17:33:05,185][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/NPO_lr2e-05_beta0.05_alpha1_epoch5/TOFU_EVAL.json
+[2026-07-22 17:33:05,185][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/NPO_lr2e-05_beta0.05_alpha1_epoch5/TOFU_SUMMARY.json
+[2026-07-22 17:33:05,981][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:05,  2.33it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:03,  3.22it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  3.58it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.83it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.03it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.12it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.20it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:02<00:01,  4.29it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.32it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.35it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.36it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.36it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.57it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.13it/s]
+Evaluated 400 examples
+[2026-07-22 17:33:09,548][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:13,  1.11s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:12,  1.11s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:11,  1.14s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:10,  1.15s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:09,  1.13s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.13s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.13s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.11s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:10<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:11<00:03,  1.10s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:12<00:02,  1.11s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.12s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.06s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.11s/it]
+Evaluated 400 examples
+[2026-07-22 17:33:23,929][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-22 17:33:23,930][evaluator][INFO] - Result for metric forget_truth_ratio:	0.6079082560719065
+[2026-07-22 17:33:23,935][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-22 17:33:23,935][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:33:23,943][metrics][INFO] - Evaluating forget_quality
+[2026-07-22 17:33:23,944][evaluator][INFO] - Result for metric forget_quality:	0.010032513978788836
+[2026-07-22 17:33:24,365][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.46it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.32it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.19it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:02,  4.21it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.27it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.35it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.37it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.48it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.54it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.54it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.53it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.52it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.75it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.49it/s]
+Evaluated 400 examples
+[2026-07-22 17:33:27,260][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.027278467593714596
+[2026-07-22 17:33:27,563][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:03,  3.13it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:03,  3.24it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:03,  2.97it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:02,  3.07it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:02,  2.88it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:02<00:02,  2.89it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:02<00:02,  2.25it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  2.60it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:03<00:01,  2.78it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:03<00:01,  2.87it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:03<00:00,  2.99it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:04<00:00,  2.88it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:04<00:00,  2.93it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:04<00:00,  2.85it/s]
+Evaluated 400 examples
+[2026-07-22 17:33:32,131][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.23939069192587226
+[2026-07-22 17:33:32,433][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.63it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.68it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.67it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.62it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.61it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.53it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.54it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.54it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.60it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.61it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.56it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.61it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.84it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.65it/s]
+Evaluated 400 examples
+[2026-07-22 17:33:35,616][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:03,  3.42it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:03,  3.58it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:03,  3.26it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:02,  3.37it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:02,  3.18it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:01<00:02,  3.36it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:02<00:01,  3.31it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  3.16it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:02<00:01,  3.36it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:03<00:00,  3.31it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:03<00:00,  3.38it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:03<00:00,  3.35it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.50it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.36it/s]
+Evaluated 400 examples
+[2026-07-22 17:33:39,799][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.46it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.48it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.50it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.55it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.59it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.54it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.47it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.40it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.39it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.42it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.42it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.46it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.71it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.53it/s]
+Evaluated 400 examples
+[2026-07-22 17:33:42,991][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:12,  1.06s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:11,  1.06s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:10,  1.08s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:09,  1.07s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:08,  1.06s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.07s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.09s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.10s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:09<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:10<00:03,  1.09s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:11<00:02,  1.10s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.10s/it]
Calculating loss: 100%|██████████| 13/13 [00:13<00:00,  1.05s/it]
Calculating loss: 100%|██████████| 13/13 [00:13<00:00,  1.08s/it]
+Evaluated 400 examples
+[2026-07-22 17:33:56,980][metrics][INFO] - Evaluating retain_Truth_Ratio
+[2026-07-22 17:33:57,434][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.01it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.35it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  6.41it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  6.59it/s]
+Evaluated 100 examples
+[2026-07-22 17:33:58,414][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.41it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.50it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.10it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.51it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.54it/s]
+Evaluated 100 examples
+[2026-07-22 17:33:59,989][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+[2026-07-22 17:34:00,325][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:01,  2.31it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:01<00:01,  1.90it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:01<00:00,  2.03it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  2.88it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  2.50it/s]
+Evaluated 100 examples
+[2026-07-22 17:34:01,929][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-22 17:34:01,929][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:34:01,929][metrics][INFO] - Evaluating ra_Truth_Ratio
+[2026-07-22 17:34:02,221][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  7.97it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.54it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  8.29it/s]
+Evaluated 117 examples
+[2026-07-22 17:34:03,002][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.66it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.48it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.09it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.56it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.57it/s]
+Evaluated 117 examples
+[2026-07-22 17:34:04,558][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+[2026-07-22 17:34:04,925][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  3.26it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  2.95it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  3.22it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  3.44it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  3.32it/s]
+Evaluated 117 examples
+[2026-07-22 17:34:06,131][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-22 17:34:06,132][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:34:06,132][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-22 17:34:06,132][metrics][INFO] - Evaluating model_utility
+[2026-07-22 17:34:06,132][evaluator][INFO] - Result for metric model_utility:	0.5027053232639895
+[2026-07-22 17:34:06,833][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:34:06,840][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.94it/s]
 31%|███       | 4/13 [00:00<00:00, 12.73it/s]
 46%|████▌     | 6/13 [00:00<00:00, 13.44it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.84it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 14.16it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.91it/s]
100%|██████████| 13/13 [00:00<00:00, 14.37it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.75it/s]
 31%|███       | 4/13 [00:00<00:00, 13.59it/s]
 46%|████▌     | 6/13 [00:00<00:00, 13.99it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 14.01it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 14.06it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 14.16it/s]
100%|██████████| 13/13 [00:00<00:00, 14.51it/s]
+[2026-07-22 17:34:08,643][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:34:08,649][metrics][INFO] - Evaluating privleak
+[2026-07-22 17:34:08,649][evaluator][INFO] - Result for metric privleak:	56.96635013553622
+[2026-07-22 17:34:08,962][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:00, 11.57it/s]
Calculating ES:  31%|███       | 4/13 [00:00<00:00, 10.61it/s]
Calculating ES:  46%|████▌     | 6/13 [00:00<00:00, 11.29it/s]
Calculating ES:  62%|██████▏   | 8/13 [00:00<00:00, 11.75it/s]
Calculating ES:  77%|███████▋  | 10/13 [00:00<00:00, 11.89it/s]
Calculating ES:  92%|█████████▏| 12/13 [00:01<00:00, 11.54it/s]
Calculating ES: 100%|██████████| 13/13 [00:01<00:00, 11.98it/s]
+Evaluated 400 examples
+[2026-07-22 17:34:10,048][evaluator][INFO] - Result for metric extraction_strength:	0.0700777811566826
diff --git a/logs/eval_NPO_lr2e-05_beta0.05_alpha2_epoch10.log b/logs/eval_NPO_lr2e-05_beta0.05_alpha2_epoch10.log
new file mode 100644
index 0000000000000000000000000000000000000000..b0f00a699e675d1d266439848107a409c8439d5a
--- /dev/null
+++ b/logs/eval_NPO_lr2e-05_beta0.05_alpha2_epoch10.log
@@ -0,0 +1,116 @@
+[2026-07-21 07:51:40,114][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/NPO_lr2e-05_beta0.05_alpha2_epoch10
+[2026-07-21 07:51:40,115][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-21 07:51:40,115][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/NPO_lr2e-05_beta0.05_alpha2_epoch10/TOFU_EVAL.json
+[2026-07-21 07:51:40,116][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/NPO_lr2e-05_beta0.05_alpha2_epoch10/TOFU_SUMMARY.json
+[2026-07-21 07:51:40,749][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:07,  1.51it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:04,  2.52it/s]
Calculating loss:  23%|██▎       | 3/13 [00:01<00:03,  3.14it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.60it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:02,  3.95it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.19it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.37it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:02<00:01,  4.53it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.60it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.66it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.67it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.68it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.83it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.09it/s]
+Evaluated 400 examples
+[2026-07-21 07:51:44,360][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:12,  1.03s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:11,  1.02s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:10,  1.04s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:09,  1.04s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:08,  1.03s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.03s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.03s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.01s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:09<00:04,  1.01s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:10<00:03,  1.01s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:11<00:02,  1.02s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:12<00:01,  1.02s/it]
Calculating loss: 100%|██████████| 13/13 [00:13<00:00,  1.03it/s]
Calculating loss: 100%|██████████| 13/13 [00:13<00:00,  1.01s/it]
+Evaluated 400 examples
+[2026-07-21 07:51:57,492][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-21 07:51:57,493][evaluator][INFO] - Result for metric forget_truth_ratio:	0.5738164572140625
+[2026-07-21 07:51:57,502][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-21 07:51:57,503][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-21 07:51:57,518][metrics][INFO] - Evaluating forget_quality
+[2026-07-21 07:51:57,520][evaluator][INFO] - Result for metric forget_quality:	0.010032513978788836
+[2026-07-21 07:51:57,889][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.81it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.81it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.68it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.71it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.74it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.78it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.78it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.86it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.89it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.87it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.85it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.84it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  5.09it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.88it/s]
+Evaluated 400 examples
+[2026-07-21 07:52:00,556][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.01426412469943898
+[2026-07-21 07:52:00,952][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:07,  1.64it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:01<00:08,  1.26it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:02<00:07,  1.41it/s]
Calculating text similarity:  31%|███       | 4/13 [00:02<00:06,  1.41it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:03<00:05,  1.56it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:03<00:04,  1.72it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:04<00:03,  1.74it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:04<00:02,  1.73it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:05<00:02,  1.77it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:06<00:01,  1.81it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:06<00:01,  1.91it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:07<00:00,  1.83it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:07<00:00,  1.86it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:07<00:00,  1.71it/s]
+Evaluated 400 examples
+[2026-07-21 07:52:08,573][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.2484264780512099
+[2026-07-21 07:52:09,002][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.95it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  5.02it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:01,  5.04it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  5.05it/s]
Calculating loss:  38%|███▊      | 5/13 [00:00<00:01,  5.05it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.98it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.97it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.94it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.97it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.96it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.91it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.95it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  5.19it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  5.03it/s]
+Evaluated 400 examples
+[2026-07-21 07:52:11,924][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:06,  1.72it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:01<00:05,  2.01it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:01<00:05,  1.80it/s]
Calculating text similarity:  31%|███       | 4/13 [00:02<00:05,  1.70it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:02<00:04,  1.63it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:03<00:04,  1.60it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:04<00:03,  1.52it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:05<00:03,  1.39it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:05<00:02,  1.47it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:06<00:01,  1.55it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:07<00:01,  1.48it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:07<00:00,  1.56it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:08<00:00,  1.61it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:08<00:00,  1.58it/s]
+Evaluated 400 examples
+[2026-07-21 07:52:20,491][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.91it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.93it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.90it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.96it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.98it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.91it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.88it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.83it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.82it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.84it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.82it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.84it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  5.09it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.93it/s]
+Using the latest cached version of the dataset since locuslab/TOFU couldn't be found on the Hugging Face Hub
+Evaluated 400 examples
+[2026-07-21 07:52:23,165][datasets.load][WARNING] - Using the latest cached version of the dataset since locuslab/TOFU couldn't be found on the Hugging Face Hub
+Found the latest cached dataset configuration 'retain_perturbed' at /workspace/hf_cache_ou/locuslab___tofu/retain_perturbed/0.0.0/324592d84ae4f482ac7249b9285c2ecdb53e3a68 (last modified on Mon Jul 20 13:22:39 2026).
+[2026-07-21 07:52:23,166][datasets.packaged_modules.cache.cache][WARNING] - Found the latest cached dataset configuration 'retain_perturbed' at /workspace/hf_cache_ou/locuslab___tofu/retain_perturbed/0.0.0/324592d84ae4f482ac7249b9285c2ecdb53e3a68 (last modified on Mon Jul 20 13:22:39 2026).
+[2026-07-21 07:52:23,172][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:11,  1.02it/s]
Calculating loss:  15%|█▌        | 2/13 [00:01<00:10,  1.02it/s]
Calculating loss:  23%|██▎       | 3/13 [00:02<00:09,  1.01it/s]
Calculating loss:  31%|███       | 4/13 [00:03<00:08,  1.02it/s]
Calculating loss:  38%|███▊      | 5/13 [00:04<00:07,  1.03it/s]
Calculating loss:  46%|████▌     | 6/13 [00:05<00:06,  1.01it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:06<00:05,  1.00it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:07<00:05,  1.00s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:08<00:04,  1.00s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:09<00:03,  1.00s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:10<00:02,  1.01s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:11<00:01,  1.01s/it]
Calculating loss: 100%|██████████| 13/13 [00:12<00:00,  1.04it/s]
Calculating loss: 100%|██████████| 13/13 [00:12<00:00,  1.01it/s]
+Evaluated 400 examples
+[2026-07-21 07:52:35,989][metrics][INFO] - Evaluating retain_Truth_Ratio
+Using the latest cached version of the dataset since locuslab/TOFU couldn't be found on the Hugging Face Hub
+[2026-07-21 07:52:36,028][datasets.load][WARNING] - Using the latest cached version of the dataset since locuslab/TOFU couldn't be found on the Hugging Face Hub
+Found the latest cached dataset configuration 'real_authors_perturbed' at /workspace/hf_cache_ou/locuslab___tofu/real_authors_perturbed/0.0.0/324592d84ae4f482ac7249b9285c2ecdb53e3a68 (last modified on Mon Jul 20 13:23:07 2026).
+[2026-07-21 07:52:36,029][datasets.packaged_modules.cache.cache][WARNING] - Found the latest cached dataset configuration 'real_authors_perturbed' at /workspace/hf_cache_ou/locuslab___tofu/real_authors_perturbed/0.0.0/324592d84ae4f482ac7249b9285c2ecdb53e3a68 (last modified on Mon Jul 20 13:23:07 2026).
+[2026-07-21 07:52:36,034][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.48it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.88it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  8.01it/s]
+Using the latest cached version of the dataset since locuslab/TOFU couldn't be found on the Hugging Face Hub
+Evaluated 100 examples
+[2026-07-21 07:52:36,569][datasets.load][WARNING] - Using the latest cached version of the dataset since locuslab/TOFU couldn't be found on the Hugging Face Hub
+Found the latest cached dataset configuration 'real_authors_perturbed' at /workspace/hf_cache_ou/locuslab___tofu/real_authors_perturbed/0.0.0/324592d84ae4f482ac7249b9285c2ecdb53e3a68 (last modified on Mon Jul 20 13:23:07 2026).
+[2026-07-21 07:52:36,571][datasets.packaged_modules.cache.cache][WARNING] - Found the latest cached dataset configuration 'real_authors_perturbed' at /workspace/hf_cache_ou/locuslab___tofu/real_authors_perturbed/0.0.0/324592d84ae4f482ac7249b9285c2ecdb53e3a68 (last modified on Mon Jul 20 13:23:07 2026).
+[2026-07-21 07:52:36,575][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.38it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.63it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.22it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.64it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.67it/s]
+Evaluated 100 examples
+[2026-07-21 07:52:38,073][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+Using the latest cached version of the dataset since locuslab/TOFU couldn't be found on the Hugging Face Hub
+[2026-07-21 07:52:38,110][datasets.load][WARNING] - Using the latest cached version of the dataset since locuslab/TOFU couldn't be found on the Hugging Face Hub
+Found the latest cached dataset configuration 'real_authors_perturbed' at /workspace/hf_cache_ou/locuslab___tofu/real_authors_perturbed/0.0.0/324592d84ae4f482ac7249b9285c2ecdb53e3a68 (last modified on Mon Jul 20 13:23:07 2026).
+[2026-07-21 07:52:38,112][datasets.packaged_modules.cache.cache][WARNING] - Found the latest cached dataset configuration 'real_authors_perturbed' at /workspace/hf_cache_ou/locuslab___tofu/real_authors_perturbed/0.0.0/324592d84ae4f482ac7249b9285c2ecdb53e3a68 (last modified on Mon Jul 20 13:23:07 2026).
+[2026-07-21 07:52:38,117][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:01,  2.32it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  2.08it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:01<00:00,  2.39it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  3.00it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  2.68it/s]
+Evaluated 100 examples
+[2026-07-21 07:52:39,611][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-21 07:52:39,612][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-21 07:52:39,612][metrics][INFO] - Evaluating ra_Truth_Ratio
+Using the latest cached version of the dataset since locuslab/TOFU couldn't be found on the Hugging Face Hub
+[2026-07-21 07:52:39,649][datasets.load][WARNING] - Using the latest cached version of the dataset since locuslab/TOFU couldn't be found on the Hugging Face Hub
+Found the latest cached dataset configuration 'world_facts_perturbed' at /workspace/hf_cache_ou/locuslab___tofu/world_facts_perturbed/0.0.0/324592d84ae4f482ac7249b9285c2ecdb53e3a68 (last modified on Mon Jul 20 13:23:12 2026).
+[2026-07-21 07:52:39,651][datasets.packaged_modules.cache.cache][WARNING] - Found the latest cached dataset configuration 'world_facts_perturbed' at /workspace/hf_cache_ou/locuslab___tofu/world_facts_perturbed/0.0.0/324592d84ae4f482ac7249b9285c2ecdb53e3a68 (last modified on Mon Jul 20 13:23:12 2026).
+[2026-07-21 07:52:39,655][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.39it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.90it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  8.68it/s]
+Using the latest cached version of the dataset since locuslab/TOFU couldn't be found on the Hugging Face Hub
+Evaluated 117 examples
+[2026-07-21 07:52:40,152][datasets.load][WARNING] - Using the latest cached version of the dataset since locuslab/TOFU couldn't be found on the Hugging Face Hub
+Found the latest cached dataset configuration 'world_facts_perturbed' at /workspace/hf_cache_ou/locuslab___tofu/world_facts_perturbed/0.0.0/324592d84ae4f482ac7249b9285c2ecdb53e3a68 (last modified on Mon Jul 20 13:23:12 2026).
+[2026-07-21 07:52:40,153][datasets.packaged_modules.cache.cache][WARNING] - Found the latest cached dataset configuration 'world_facts_perturbed' at /workspace/hf_cache_ou/locuslab___tofu/world_facts_perturbed/0.0.0/324592d84ae4f482ac7249b9285c2ecdb53e3a68 (last modified on Mon Jul 20 13:23:12 2026).
+[2026-07-21 07:52:40,157][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.56it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.62it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.22it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.71it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.72it/s]
+Evaluated 117 examples
+[2026-07-21 07:52:41,629][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+Using the latest cached version of the dataset since locuslab/TOFU couldn't be found on the Hugging Face Hub
+[2026-07-21 07:52:41,664][datasets.load][WARNING] - Using the latest cached version of the dataset since locuslab/TOFU couldn't be found on the Hugging Face Hub
+Found the latest cached dataset configuration 'world_facts_perturbed' at /workspace/hf_cache_ou/locuslab___tofu/world_facts_perturbed/0.0.0/324592d84ae4f482ac7249b9285c2ecdb53e3a68 (last modified on Mon Jul 20 13:23:12 2026).
+[2026-07-21 07:52:41,667][datasets.packaged_modules.cache.cache][WARNING] - Found the latest cached dataset configuration 'world_facts_perturbed' at /workspace/hf_cache_ou/locuslab___tofu/world_facts_perturbed/0.0.0/324592d84ae4f482ac7249b9285c2ecdb53e3a68 (last modified on Mon Jul 20 13:23:12 2026).
+[2026-07-21 07:52:41,671][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:01,  1.97it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  2.13it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:02<00:00,  1.31it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:02<00:00,  1.52it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:02<00:00,  1.56it/s]
+Evaluated 117 examples
+[2026-07-21 07:52:44,228][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-21 07:52:44,228][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-21 07:52:44,229][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-21 07:52:44,229][metrics][INFO] - Evaluating model_utility
+[2026-07-21 07:52:44,230][evaluator][INFO] - Result for metric model_utility:	0.5789363789267417
+Using the latest cached version of the dataset since locuslab/TOFU couldn't be found on the Hugging Face Hub
+[2026-07-21 07:52:44,328][datasets.load][WARNING] - Using the latest cached version of the dataset since locuslab/TOFU couldn't be found on the Hugging Face Hub
+Found the latest cached dataset configuration 'forget10_perturbed' at /workspace/hf_cache_ou/locuslab___tofu/forget10_perturbed/0.0.0/324592d84ae4f482ac7249b9285c2ecdb53e3a68 (last modified on Mon Jul 20 13:17:15 2026).
+[2026-07-21 07:52:44,329][datasets.packaged_modules.cache.cache][WARNING] - Found the latest cached dataset configuration 'forget10_perturbed' at /workspace/hf_cache_ou/locuslab___tofu/forget10_perturbed/0.0.0/324592d84ae4f482ac7249b9285c2ecdb53e3a68 (last modified on Mon Jul 20 13:17:15 2026).
+Using the latest cached version of the dataset since locuslab/TOFU couldn't be found on the Hugging Face Hub
+[2026-07-21 07:52:44,369][datasets.load][WARNING] - Using the latest cached version of the dataset since locuslab/TOFU couldn't be found on the Hugging Face Hub
+Found the latest cached dataset configuration 'holdout10' at /workspace/hf_cache_ou/locuslab___tofu/holdout10/0.0.0/324592d84ae4f482ac7249b9285c2ecdb53e3a68 (last modified on Mon Jul 20 13:23:17 2026).
+[2026-07-21 07:52:44,372][datasets.packaged_modules.cache.cache][WARNING] - Found the latest cached dataset configuration 'holdout10' at /workspace/hf_cache_ou/locuslab___tofu/holdout10/0.0.0/324592d84ae4f482ac7249b9285c2ecdb53e3a68 (last modified on Mon Jul 20 13:23:17 2026).
+[2026-07-21 07:52:44,376][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-21 07:52:44,392][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.45it/s]
 31%|███       | 4/13 [00:00<00:00, 11.50it/s]
 46%|████▌     | 6/13 [00:00<00:00, 11.77it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 11.94it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 12.17it/s]
 92%|█████████▏| 12/13 [00:01<00:00, 11.91it/s]
100%|██████████| 13/13 [00:01<00:00, 12.47it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 11.66it/s]
 31%|███       | 4/13 [00:00<00:00, 11.53it/s]
 46%|████▌     | 6/13 [00:00<00:00, 11.86it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 11.89it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 11.91it/s]
 92%|█████████▏| 12/13 [00:01<00:00, 12.02it/s]
100%|██████████| 13/13 [00:01<00:00, 12.32it/s]
+[2026-07-21 07:52:46,494][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-21 07:52:46,507][metrics][INFO] - Evaluating privleak
+[2026-07-21 07:52:46,508][evaluator][INFO] - Result for metric privleak:	58.33552670594473
+Using the latest cached version of the dataset since locuslab/TOFU couldn't be found on the Hugging Face Hub
+[2026-07-21 07:52:46,581][datasets.load][WARNING] - Using the latest cached version of the dataset since locuslab/TOFU couldn't be found on the Hugging Face Hub
+Found the latest cached dataset configuration 'forget10_perturbed' at /workspace/hf_cache_ou/locuslab___tofu/forget10_perturbed/0.0.0/324592d84ae4f482ac7249b9285c2ecdb53e3a68 (last modified on Mon Jul 20 13:17:15 2026).
+[2026-07-21 07:52:46,583][datasets.packaged_modules.cache.cache][WARNING] - Found the latest cached dataset configuration 'forget10_perturbed' at /workspace/hf_cache_ou/locuslab___tofu/forget10_perturbed/0.0.0/324592d84ae4f482ac7249b9285c2ecdb53e3a68 (last modified on Mon Jul 20 13:17:15 2026).
+[2026-07-21 07:52:46,587][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:01, 10.42it/s]
Calculating ES:  31%|███       | 4/13 [00:00<00:00,  9.38it/s]
Calculating ES:  46%|████▌     | 6/13 [00:00<00:00,  9.78it/s]
Calculating ES:  54%|█████▍    | 7/13 [00:00<00:00,  9.78it/s]
Calculating ES:  69%|██████▉   | 9/13 [00:00<00:00, 10.23it/s]
Calculating ES:  85%|████████▍ | 11/13 [00:01<00:00, 10.09it/s]
Calculating ES: 100%|██████████| 13/13 [00:01<00:00, 10.88it/s]
Calculating ES: 100%|██████████| 13/13 [00:01<00:00, 10.33it/s]
+Evaluated 400 examples
+[2026-07-21 07:52:47,846][evaluator][INFO] - Result for metric extraction_strength:	0.09007475310833865
diff --git a/logs/eval_NPO_lr2e-05_beta0.05_alpha2_epoch5.log b/logs/eval_NPO_lr2e-05_beta0.05_alpha2_epoch5.log
new file mode 100644
index 0000000000000000000000000000000000000000..390f83edda265104caa28014d23cceeb37d062fb
--- /dev/null
+++ b/logs/eval_NPO_lr2e-05_beta0.05_alpha2_epoch5.log
@@ -0,0 +1,78 @@
+/opt/conda/envs/copygen/lib/python3.12/site-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.6.3) or chardet (6.0.0.post1)/charset_normalizer (3.4.4) doesn't match a supported version!
+  warnings.warn(
+[2026-07-22 17:28:52,242][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/NPO_lr2e-05_beta0.05_alpha2_epoch5
+[2026-07-22 17:28:52,243][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-22 17:28:52,243][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/NPO_lr2e-05_beta0.05_alpha2_epoch5/TOFU_EVAL.json
+[2026-07-22 17:28:52,243][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/NPO_lr2e-05_beta0.05_alpha2_epoch5/TOFU_SUMMARY.json
+[2026-07-22 17:28:53,064][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:04,  2.47it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:03,  3.33it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  3.68it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.91it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.06it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.15it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.24it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:02<00:01,  4.37it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.40it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.43it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.41it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.38it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.59it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.19it/s]
+Evaluated 400 examples
+[2026-07-22 17:28:56,601][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:13,  1.12s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:12,  1.13s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:11,  1.15s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:10,  1.15s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:09,  1.13s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.13s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.12s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.11s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:10<00:04,  1.11s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:11<00:03,  1.10s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:12<00:02,  1.11s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.11s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.06s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.11s/it]
+Evaluated 400 examples
+[2026-07-22 17:29:10,978][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-22 17:29:10,979][evaluator][INFO] - Result for metric forget_truth_ratio:	0.6066953898933433
+[2026-07-22 17:29:10,983][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-22 17:29:10,983][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:29:10,991][metrics][INFO] - Evaluating forget_quality
+[2026-07-22 17:29:10,991][evaluator][INFO] - Result for metric forget_quality:	0.11114884622967085
+[2026-07-22 17:29:11,469][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.45it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.41it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.28it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:02,  4.32it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.37it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.42it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.40it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.50it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.55it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.53it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.51it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.43it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.66it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.49it/s]
+Evaluated 400 examples
+[2026-07-22 17:29:14,365][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.025932830879464745
+[2026-07-22 17:29:14,686][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:04,  2.93it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:03,  3.47it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:01<00:03,  2.73it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:03,  2.78it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:02,  2.95it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:01<00:02,  3.33it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:02<00:01,  3.77it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  3.92it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:02<00:01,  3.95it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:02<00:00,  3.69it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:03<00:00,  3.41it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:03<00:00,  3.35it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.18it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.32it/s]
+Evaluated 400 examples
+[2026-07-22 17:29:18,606][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.23124368132458442
+[2026-07-22 17:29:18,922][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.65it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.68it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.69it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.71it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.72it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.66it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.64it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.61it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.64it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.63it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.58it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.62it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.84it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.69it/s]
+Evaluated 400 examples
+[2026-07-22 17:29:22,025][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:03,  3.20it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:03,  3.29it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:03,  3.27it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:02,  3.54it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:02,  3.65it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:01<00:01,  3.63it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:02<00:01,  3.37it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  3.35it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:02<00:01,  3.17it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:02<00:00,  3.33it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:03<00:00,  3.31it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:03<00:00,  3.33it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.36it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.37it/s]
+Evaluated 400 examples
+[2026-07-22 17:29:26,192][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.58it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.61it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.57it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.63it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.66it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.58it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.54it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.49it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.49it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.50it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.48it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.50it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.74it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.59it/s]
+Evaluated 400 examples
+[2026-07-22 17:29:29,330][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:12,  1.07s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:11,  1.08s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:10,  1.08s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:09,  1.07s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:08,  1.06s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.08s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.09s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.10s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:09<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:10<00:03,  1.10s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:11<00:02,  1.11s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.10s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.05s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.08s/it]
+Evaluated 400 examples
+[2026-07-22 17:29:43,347][metrics][INFO] - Evaluating retain_Truth_Ratio
+[2026-07-22 17:29:45,162][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.09it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.52it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  7.48it/s]
+Evaluated 100 examples
+[2026-07-22 17:29:46,078][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.32it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.49it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.06it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.06it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.22it/s]
+Evaluated 100 examples
+[2026-07-22 17:29:47,881][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+[2026-07-22 17:29:48,234][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  3.60it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  3.22it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  3.21it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  4.36it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  3.89it/s]
+Evaluated 100 examples
+[2026-07-22 17:29:49,263][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-22 17:29:49,263][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:29:49,263][metrics][INFO] - Evaluating ra_Truth_Ratio
+[2026-07-22 17:29:49,704][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.06it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.57it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  8.31it/s]
+Evaluated 117 examples
+[2026-07-22 17:29:50,551][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.74it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.48it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.10it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.54it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.56it/s]
+Evaluated 117 examples
+[2026-07-22 17:29:52,112][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+[2026-07-22 17:29:52,413][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  4.55it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  4.78it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  3.67it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  3.44it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  3.67it/s]
+Evaluated 117 examples
+[2026-07-22 17:29:53,503][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-22 17:29:53,503][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:29:53,503][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-22 17:29:53,503][metrics][INFO] - Evaluating model_utility
+[2026-07-22 17:29:53,504][evaluator][INFO] - Result for metric model_utility:	0.540957951431643
+[2026-07-22 17:29:54,191][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:29:54,202][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.00it/s]
 31%|███       | 4/13 [00:00<00:00, 11.88it/s]
 46%|████▌     | 6/13 [00:00<00:00, 12.36it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 12.85it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.26it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.01it/s]
100%|██████████| 13/13 [00:00<00:00, 13.36it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 12.51it/s]
 31%|███       | 4/13 [00:00<00:00, 12.45it/s]
 46%|████▌     | 6/13 [00:00<00:00, 12.98it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 12.74it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 12.81it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 12.91it/s]
100%|██████████| 13/13 [00:00<00:00, 13.25it/s]
+[2026-07-22 17:29:56,160][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:29:56,167][metrics][INFO] - Evaluating privleak
+[2026-07-22 17:29:56,167][evaluator][INFO] - Result for metric privleak:	56.648984623652254
+[2026-07-22 17:29:56,489][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:00, 11.36it/s]
Calculating ES:  31%|███       | 4/13 [00:00<00:00, 10.38it/s]
Calculating ES:  46%|████▌     | 6/13 [00:00<00:00, 10.96it/s]
Calculating ES:  62%|██████▏   | 8/13 [00:00<00:00, 11.42it/s]
Calculating ES:  77%|███████▋  | 10/13 [00:00<00:00, 11.59it/s]
Calculating ES:  92%|█████████▏| 12/13 [00:01<00:00, 11.42it/s]
Calculating ES: 100%|██████████| 13/13 [00:01<00:00, 11.77it/s]
+Evaluated 400 examples
+[2026-07-22 17:29:57,595][evaluator][INFO] - Result for metric extraction_strength:	0.07766172367731558
diff --git a/logs/eval_NPO_lr2e-05_beta0.05_alpha5_epoch10.log b/logs/eval_NPO_lr2e-05_beta0.05_alpha5_epoch10.log
new file mode 100644
index 0000000000000000000000000000000000000000..7d1d71fee63f204320957e0382410f3df00aa109
--- /dev/null
+++ b/logs/eval_NPO_lr2e-05_beta0.05_alpha5_epoch10.log
@@ -0,0 +1,78 @@
+/opt/conda/envs/copygen/lib/python3.12/site-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.6.3) or chardet (6.0.0.post1)/charset_normalizer (3.4.4) doesn't match a supported version!
+  warnings.warn(
+[2026-07-22 17:26:48,547][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/NPO_lr2e-05_beta0.05_alpha5_epoch10
+[2026-07-22 17:26:48,548][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-22 17:26:48,548][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/NPO_lr2e-05_beta0.05_alpha5_epoch10/TOFU_EVAL.json
+[2026-07-22 17:26:48,548][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/NPO_lr2e-05_beta0.05_alpha5_epoch10/TOFU_SUMMARY.json
+[2026-07-22 17:26:49,343][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:05,  2.32it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:03,  3.21it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  3.57it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.82it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.02it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.12it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.21it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:02<00:01,  4.32it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.36it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.40it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.39it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.37it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.56it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.14it/s]
+Evaluated 400 examples
+[2026-07-22 17:26:52,915][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:13,  1.13s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:12,  1.12s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:11,  1.15s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:10,  1.15s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:09,  1.14s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.13s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.13s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:09<00:05,  1.12s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:10<00:04,  1.11s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:11<00:03,  1.11s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:12<00:02,  1.11s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.12s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.06s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.11s/it]
+Evaluated 400 examples
+[2026-07-22 17:27:07,329][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-22 17:27:07,329][evaluator][INFO] - Result for metric forget_truth_ratio:	0.612589911270902
+[2026-07-22 17:27:07,334][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-22 17:27:07,335][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:27:07,342][metrics][INFO] - Evaluating forget_quality
+[2026-07-22 17:27:07,343][evaluator][INFO] - Result for metric forget_quality:	0.7582774273046589
+[2026-07-22 17:27:07,852][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.47it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.45it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.29it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:02,  4.32it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.34it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.40it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.34it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.40it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.44it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.42it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.42it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.43it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.68it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.46it/s]
+Evaluated 400 examples
+[2026-07-22 17:27:10,767][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.029849014952778818
+[2026-07-22 17:27:11,377][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:03,  3.13it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:03,  2.84it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:01<00:03,  2.73it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:03,  2.77it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:02,  2.90it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:02<00:02,  2.84it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:02<00:02,  2.67it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  2.80it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:03<00:01,  3.13it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:03<00:00,  3.08it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:03<00:00,  3.03it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:04<00:00,  3.09it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:04<00:00,  3.24it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:04<00:00,  2.99it/s]
+Evaluated 400 examples
+[2026-07-22 17:27:15,730][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.22600664040140728
+[2026-07-22 17:27:16,040][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.56it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.65it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.61it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.66it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.68it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.62it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.62it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.51it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.53it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.52it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.49it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.55it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.79it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.63it/s]
+Evaluated 400 examples
+[2026-07-22 17:27:19,148][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:03,  3.59it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:03,  3.25it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:02,  3.40it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:02,  3.39it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:02,  3.31it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:01<00:02,  2.92it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:02<00:02,  2.49it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  2.59it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:03<00:01,  2.73it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:03<00:01,  2.74it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:03<00:00,  2.81it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:04<00:00,  2.87it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:04<00:00,  3.05it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:04<00:00,  2.94it/s]
+Evaluated 400 examples
+[2026-07-22 17:27:23,889][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.50it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.56it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.54it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.59it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.61it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.53it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.51it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.45it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.43it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.41it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.37it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.39it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.65it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.52it/s]
+Evaluated 400 examples
+[2026-07-22 17:27:27,062][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:12,  1.06s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:11,  1.06s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:10,  1.07s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:09,  1.07s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:08,  1.06s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.08s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.10s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.10s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:09<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:10<00:03,  1.10s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:11<00:02,  1.11s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.11s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.05s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.08s/it]
+Evaluated 400 examples
+[2026-07-22 17:27:41,095][metrics][INFO] - Evaluating retain_Truth_Ratio
+[2026-07-22 17:27:41,517][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  7.96it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.48it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  7.46it/s]
+Evaluated 100 examples
+[2026-07-22 17:27:42,370][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.17it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.44it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.07it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.47it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.50it/s]
+Evaluated 100 examples
+[2026-07-22 17:27:43,973][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+[2026-07-22 17:27:44,294][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  3.46it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  4.24it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  4.35it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  5.51it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  4.89it/s]
+Evaluated 100 examples
+[2026-07-22 17:27:45,113][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-22 17:27:45,113][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:27:45,113][metrics][INFO] - Evaluating ra_Truth_Ratio
+[2026-07-22 17:27:45,456][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.17it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.64it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  8.43it/s]
+Evaluated 117 examples
+[2026-07-22 17:27:46,281][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.84it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.49it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.09it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.56it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.57it/s]
+Evaluated 117 examples
+[2026-07-22 17:27:47,837][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+[2026-07-22 17:27:48,143][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  4.26it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  4.77it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  3.75it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  3.96it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  4.03it/s]
+Evaluated 117 examples
+[2026-07-22 17:27:49,136][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-22 17:27:49,136][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:27:49,136][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-22 17:27:49,137][metrics][INFO] - Evaluating model_utility
+[2026-07-22 17:27:49,137][evaluator][INFO] - Result for metric model_utility:	0.5705193260103165
+[2026-07-22 17:27:49,803][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:27:49,810][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.44it/s]
 31%|███       | 4/13 [00:00<00:00, 12.31it/s]
 46%|████▌     | 6/13 [00:00<00:00, 12.62it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 12.99it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.32it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.04it/s]
100%|██████████| 13/13 [00:00<00:00, 13.53it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 12.83it/s]
 31%|███       | 4/13 [00:00<00:00, 12.68it/s]
 46%|████▌     | 6/13 [00:00<00:00, 13.10it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.22it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.04it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.19it/s]
100%|██████████| 13/13 [00:00<00:00, 13.56it/s]
+[2026-07-22 17:27:51,732][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:27:51,739][metrics][INFO] - Evaluating privleak
+[2026-07-22 17:27:51,739][evaluator][INFO] - Result for metric privleak:	42.34577148997254
+[2026-07-22 17:27:52,063][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:00, 11.71it/s]
Calculating ES:  31%|███       | 4/13 [00:00<00:00, 10.67it/s]
Calculating ES:  46%|████▌     | 6/13 [00:00<00:00, 11.32it/s]
Calculating ES:  62%|██████▏   | 8/13 [00:00<00:00, 11.75it/s]
Calculating ES:  77%|███████▋  | 10/13 [00:00<00:00, 11.90it/s]
Calculating ES:  92%|█████████▏| 12/13 [00:01<00:00, 11.59it/s]
Calculating ES: 100%|██████████| 13/13 [00:01<00:00, 12.00it/s]
+Evaluated 400 examples
+[2026-07-22 17:27:53,147][evaluator][INFO] - Result for metric extraction_strength:	0.0697136039856562
diff --git a/logs/eval_NPO_lr2e-05_beta0.05_alpha5_epoch5.log b/logs/eval_NPO_lr2e-05_beta0.05_alpha5_epoch5.log
new file mode 100644
index 0000000000000000000000000000000000000000..8597e4030abade248e5edea9641af506980595b7
--- /dev/null
+++ b/logs/eval_NPO_lr2e-05_beta0.05_alpha5_epoch5.log
@@ -0,0 +1,78 @@
+/opt/conda/envs/copygen/lib/python3.12/site-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.6.3) or chardet (6.0.0.post1)/charset_normalizer (3.4.4) doesn't match a supported version!
+  warnings.warn(
+[2026-07-22 17:34:48,339][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/NPO_lr2e-05_beta0.05_alpha5_epoch5
+[2026-07-22 17:34:48,340][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-22 17:34:48,340][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/NPO_lr2e-05_beta0.05_alpha5_epoch5/TOFU_EVAL.json
+[2026-07-22 17:34:48,340][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/NPO_lr2e-05_beta0.05_alpha5_epoch5/TOFU_SUMMARY.json
+[2026-07-22 17:34:49,130][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:05,  2.35it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:03,  3.22it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  3.60it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.85it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.05it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.16it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.26it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:02<00:01,  4.38it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.42it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.46it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.44it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.41it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.61it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.18it/s]
+Evaluated 400 examples
+[2026-07-22 17:34:52,655][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:13,  1.12s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:12,  1.12s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:11,  1.15s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:10,  1.15s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:09,  1.14s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.13s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.13s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.11s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:10<00:04,  1.11s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:11<00:03,  1.11s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:12<00:02,  1.12s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.13s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.07s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.11s/it]
+Evaluated 400 examples
+[2026-07-22 17:35:07,093][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-22 17:35:07,093][evaluator][INFO] - Result for metric forget_truth_ratio:	0.6290858954238209
+[2026-07-22 17:35:07,098][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-22 17:35:07,098][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:35:07,105][metrics][INFO] - Evaluating forget_quality
+[2026-07-22 17:35:07,106][evaluator][INFO] - Result for metric forget_quality:	0.9416341827445021
+[2026-07-22 17:35:07,546][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.47it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.47it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.33it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:02,  4.37it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.40it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.45it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.45it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.54it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.58it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.57it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.54it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.52it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.75it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.55it/s]
+Evaluated 400 examples
+[2026-07-22 17:35:10,407][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.046733250617980955
+[2026-07-22 17:35:10,734][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:04,  2.81it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:03,  2.78it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:01<00:04,  2.45it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:03,  2.28it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:02<00:03,  2.50it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:02<00:02,  2.69it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:02<00:02,  2.52it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:03<00:01,  2.53it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:03<00:01,  2.83it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:03<00:01,  2.72it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:04<00:00,  2.73it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:04<00:00,  2.88it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:04<00:00,  3.17it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:04<00:00,  2.75it/s]
+Evaluated 400 examples
+[2026-07-22 17:35:15,462][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.23216731478688185
+[2026-07-22 17:35:15,819][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.67it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.72it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.73it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.75it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.75it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.69it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.67it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.64it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.67it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.66it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.61it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.65it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.87it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.72it/s]
+Evaluated 400 examples
+[2026-07-22 17:35:18,936][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:09,  1.29it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:01<00:05,  1.91it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:01<00:04,  2.24it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:03,  2.53it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:02<00:03,  2.43it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:02<00:02,  2.60it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:02<00:02,  2.52it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:03<00:01,  2.61it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:03<00:01,  2.75it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:04<00:01,  2.73it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:04<00:00,  2.53it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:04<00:00,  2.78it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:05<00:00,  2.91it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:05<00:00,  2.56it/s]
+Evaluated 400 examples
+[2026-07-22 17:35:24,319][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.62it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.63it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.60it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.65it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.68it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.61it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.57it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.52it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.52it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.53it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.50it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.53it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.75it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.61it/s]
+Evaluated 400 examples
+[2026-07-22 17:35:27,441][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:12,  1.07s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:11,  1.06s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:10,  1.08s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:09,  1.08s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:08,  1.06s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.07s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.08s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.09s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:09<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:10<00:03,  1.09s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:11<00:02,  1.11s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.10s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.05s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.08s/it]
+Evaluated 400 examples
+[2026-07-22 17:35:41,447][metrics][INFO] - Evaluating retain_Truth_Ratio
+[2026-07-22 17:35:41,958][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.25it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.59it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  7.64it/s]
+Evaluated 100 examples
+[2026-07-22 17:35:42,789][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.59it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.52it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.09it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.46it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.51it/s]
+Evaluated 100 examples
+[2026-07-22 17:35:44,382][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+[2026-07-22 17:35:44,732][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  3.63it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  4.17it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  3.93it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  5.02it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  4.56it/s]
+Evaluated 100 examples
+[2026-07-22 17:35:45,610][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-22 17:35:45,611][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:35:45,611][metrics][INFO] - Evaluating ra_Truth_Ratio
+[2026-07-22 17:35:45,908][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  7.94it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.52it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  8.24it/s]
+Evaluated 117 examples
+[2026-07-22 17:35:46,719][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.62it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.44it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.05it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.52it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.53it/s]
+Evaluated 117 examples
+[2026-07-22 17:35:48,300][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+[2026-07-22 17:35:48,653][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  3.74it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  4.01it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  2.97it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  3.13it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  3.23it/s]
+Evaluated 117 examples
+[2026-07-22 17:35:49,892][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-22 17:35:49,892][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:35:49,892][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-22 17:35:49,892][metrics][INFO] - Evaluating model_utility
+[2026-07-22 17:35:49,893][evaluator][INFO] - Result for metric model_utility:	0.5545854480715288
+[2026-07-22 17:35:50,637][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:35:50,644][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 14.13it/s]
 31%|███       | 4/13 [00:00<00:00, 12.81it/s]
 46%|████▌     | 6/13 [00:00<00:00, 13.39it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.79it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 14.19it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.85it/s]
100%|██████████| 13/13 [00:00<00:00, 14.34it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.64it/s]
 31%|███       | 4/13 [00:00<00:00, 13.45it/s]
 46%|████▌     | 6/13 [00:00<00:00, 13.62it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.71it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.79it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.93it/s]
100%|██████████| 13/13 [00:00<00:00, 14.24it/s]
+[2026-07-22 17:35:52,466][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:35:52,473][metrics][INFO] - Evaluating privleak
+[2026-07-22 17:35:52,473][evaluator][INFO] - Result for metric privleak:	30.597679734267764
+[2026-07-22 17:35:52,848][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:00, 11.34it/s]
Calculating ES:  31%|███       | 4/13 [00:00<00:00, 10.30it/s]
Calculating ES:  46%|████▌     | 6/13 [00:00<00:00, 11.02it/s]
Calculating ES:  62%|██████▏   | 8/13 [00:00<00:00, 11.49it/s]
Calculating ES:  77%|███████▋  | 10/13 [00:00<00:00, 11.42it/s]
Calculating ES:  92%|█████████▏| 12/13 [00:01<00:00, 11.19it/s]
Calculating ES: 100%|██████████| 13/13 [00:01<00:00, 11.61it/s]
+Evaluated 400 examples
+[2026-07-22 17:35:53,969][evaluator][INFO] - Result for metric extraction_strength:	0.07019027151065066
diff --git a/logs/eval_NPO_lr2e-05_beta0.1_alpha1_epoch10.log b/logs/eval_NPO_lr2e-05_beta0.1_alpha1_epoch10.log
new file mode 100644
index 0000000000000000000000000000000000000000..7e0f56cb401f17102d7ee9fb30f231b7b0b8f626
--- /dev/null
+++ b/logs/eval_NPO_lr2e-05_beta0.1_alpha1_epoch10.log
@@ -0,0 +1,78 @@
+/opt/conda/envs/copygen/lib/python3.12/site-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.6.3) or chardet (6.0.0.post1)/charset_normalizer (3.4.4) doesn't match a supported version!
+  warnings.warn(
+[2026-07-20 13:27:29,186][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/NPO_lr2e-05_beta0.1_alpha1_epoch10
+[2026-07-20 13:27:29,188][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-20 13:27:29,188][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/NPO_lr2e-05_beta0.1_alpha1_epoch10/TOFU_EVAL.json
+[2026-07-20 13:27:29,188][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/NPO_lr2e-05_beta0.1_alpha1_epoch10/TOFU_SUMMARY.json
+[2026-07-20 13:27:29,850][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:08,  1.50it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:04,  2.50it/s]
Calculating loss:  23%|██▎       | 3/13 [00:01<00:03,  3.14it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.61it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:02,  3.96it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.20it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.38it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:02<00:01,  4.54it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.32it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.46it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.53it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:03<00:00,  4.57it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.75it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.03it/s]
+Evaluated 400 examples
+[2026-07-20 13:27:33,430][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:12,  1.02s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:11,  1.02s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:10,  1.04s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:09,  1.05s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:08,  1.04s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.03s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.03s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.01s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:09<00:04,  1.01s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:10<00:03,  1.01s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:11<00:02,  1.02s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:12<00:01,  1.02s/it]
Calculating loss: 100%|██████████| 13/13 [00:13<00:00,  1.03it/s]
Calculating loss: 100%|██████████| 13/13 [00:13<00:00,  1.01s/it]
+Evaluated 400 examples
+[2026-07-20 13:27:46,578][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-20 13:27:46,579][evaluator][INFO] - Result for metric forget_truth_ratio:	0.5956753070472918
+[2026-07-20 13:27:46,590][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-20 13:27:46,591][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-20 13:27:46,611][metrics][INFO] - Evaluating forget_quality
+[2026-07-20 13:27:46,613][evaluator][INFO] - Result for metric forget_quality:	0.004907932356726081
+[2026-07-20 13:27:47,026][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.81it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.80it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.68it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.71it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.74it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.78it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.76it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.85it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.88it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.86it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.84it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.82it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  5.08it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.87it/s]
+Evaluated 400 examples
+[2026-07-20 13:27:49,698][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.08635525703430176
+[2026-07-20 13:27:49,970][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:07,  1.66it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:01<00:08,  1.30it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:02<00:07,  1.26it/s]
Calculating text similarity:  31%|███       | 4/13 [00:03<00:07,  1.24it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:03<00:05,  1.34it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:04<00:05,  1.29it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:05<00:04,  1.33it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:05<00:03,  1.48it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:06<00:02,  1.52it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:07<00:01,  1.56it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:07<00:01,  1.48it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:08<00:00,  1.51it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:09<00:00,  1.56it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:09<00:00,  1.44it/s]
+Evaluated 400 examples
+[2026-07-20 13:27:58,992][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.2832882121739552
+[2026-07-20 13:27:59,401][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.96it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.94it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.98it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  5.02it/s]
Calculating loss:  38%|███▊      | 5/13 [00:00<00:01,  5.03it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.97it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.97it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.90it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.94it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.93it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.89it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.94it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  5.18it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  5.01it/s]
+Evaluated 400 examples
+[2026-07-20 13:28:02,288][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:08,  1.49it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:01<00:08,  1.23it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:02<00:06,  1.45it/s]
Calculating text similarity:  31%|███       | 4/13 [00:02<00:06,  1.45it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:03<00:05,  1.59it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:03<00:04,  1.57it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:04<00:03,  1.55it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:05<00:03,  1.49it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:06<00:02,  1.50it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:06<00:02,  1.50it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:07<00:01,  1.46it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:07<00:00,  1.55it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:08<00:00,  1.71it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:08<00:00,  1.54it/s]
+Evaluated 400 examples
+[2026-07-20 13:28:11,103][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.90it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.93it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.91it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.96it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.99it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.91it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.89it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.84it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.81it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.83it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.79it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.82it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  5.05it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.92it/s]
+Evaluated 400 examples
+[2026-07-20 13:28:14,030][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:11,  1.02it/s]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:11,  1.02s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:10,  1.04s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:09,  1.02s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:07,  1.00it/s]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.00s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.01s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.01s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:09<00:04,  1.01s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:10<00:03,  1.01s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:11<00:02,  1.01s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:12<00:01,  1.01s/it]
Calculating loss: 100%|██████████| 13/13 [00:12<00:00,  1.04it/s]
Calculating loss: 100%|██████████| 13/13 [00:12<00:00,  1.00it/s]
+Evaluated 400 examples
+[2026-07-20 13:28:27,010][metrics][INFO] - Evaluating retain_Truth_Ratio
+[2026-07-20 13:28:27,461][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.37it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.83it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  7.92it/s]
+Evaluated 100 examples
+[2026-07-20 13:28:28,225][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.26it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.63it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.22it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.64it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.67it/s]
+Evaluated 100 examples
+[2026-07-20 13:28:29,723][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+[2026-07-20 13:28:30,019][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:01,  2.89it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  2.51it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:01<00:00,  2.24it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  2.74it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  2.62it/s]
+Evaluated 100 examples
+[2026-07-20 13:28:31,544][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-20 13:28:31,545][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-20 13:28:31,545][metrics][INFO] - Evaluating ra_Truth_Ratio
+[2026-07-20 13:28:31,852][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.34it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.89it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  8.64it/s]
+Evaluated 117 examples
+[2026-07-20 13:28:32,818][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.44it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.61it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.21it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.70it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.71it/s]
+Evaluated 117 examples
+[2026-07-20 13:28:34,294][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+[2026-07-20 13:28:34,570][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:01,  1.90it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  2.13it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:01<00:00,  1.98it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:02<00:00,  1.80it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:02<00:00,  1.87it/s]
+Evaluated 117 examples
+[2026-07-20 13:28:36,711][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-20 13:28:36,711][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-20 13:28:36,711][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-20 13:28:36,712][metrics][INFO] - Evaluating model_utility
+[2026-07-20 13:28:36,712][evaluator][INFO] - Result for metric model_utility:	0.5640611574261162
+[2026-07-20 13:28:37,421][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-20 13:28:37,437][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.17it/s]
 31%|███       | 4/13 [00:00<00:00, 11.40it/s]
 46%|████▌     | 6/13 [00:00<00:00, 11.75it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 11.91it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 12.09it/s]
 92%|█████████▏| 12/13 [00:01<00:00, 11.88it/s]
100%|██████████| 13/13 [00:01<00:00, 12.41it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 11.72it/s]
 31%|███       | 4/13 [00:00<00:00, 11.56it/s]
 46%|████▌     | 6/13 [00:00<00:00, 11.89it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 11.88it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 11.86it/s]
 92%|█████████▏| 12/13 [00:01<00:00, 11.96it/s]
100%|██████████| 13/13 [00:01<00:00, 12.29it/s]
+[2026-07-20 13:28:39,547][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-20 13:28:39,562][metrics][INFO] - Evaluating privleak
+[2026-07-20 13:28:39,562][evaluator][INFO] - Result for metric privleak:	2.8001052818853487
+[2026-07-20 13:28:39,857][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:01, 10.18it/s]
Calculating ES:  31%|███       | 4/13 [00:00<00:00,  9.26it/s]
Calculating ES:  46%|████▌     | 6/13 [00:00<00:00,  9.70it/s]
Calculating ES:  54%|█████▍    | 7/13 [00:00<00:00,  9.76it/s]
Calculating ES:  69%|██████▉   | 9/13 [00:00<00:00, 10.16it/s]
Calculating ES:  85%|████████▍ | 11/13 [00:01<00:00, 10.02it/s]
Calculating ES: 100%|██████████| 13/13 [00:01<00:00, 10.80it/s]
Calculating ES: 100%|██████████| 13/13 [00:01<00:00, 10.25it/s]
+Evaluated 400 examples
+[2026-07-20 13:28:41,126][evaluator][INFO] - Result for metric extraction_strength:	0.0872873516609944
diff --git a/logs/eval_NPO_lr2e-05_beta0.1_alpha1_epoch5.log b/logs/eval_NPO_lr2e-05_beta0.1_alpha1_epoch5.log
new file mode 100644
index 0000000000000000000000000000000000000000..2e710203956d2161bfab831a15869a3a1371529d
--- /dev/null
+++ b/logs/eval_NPO_lr2e-05_beta0.1_alpha1_epoch5.log
@@ -0,0 +1,78 @@
+/opt/conda/envs/copygen/lib/python3.12/site-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.6.3) or chardet (6.0.0.post1)/charset_normalizer (3.4.4) doesn't match a supported version!
+  warnings.warn(
+[2026-07-22 17:30:28,363][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/NPO_lr2e-05_beta0.1_alpha1_epoch5
+[2026-07-22 17:30:28,364][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-22 17:30:28,364][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/NPO_lr2e-05_beta0.1_alpha1_epoch5/TOFU_EVAL.json
+[2026-07-22 17:30:28,364][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/NPO_lr2e-05_beta0.1_alpha1_epoch5/TOFU_SUMMARY.json
+[2026-07-22 17:30:29,168][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:05,  2.35it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:03,  3.19it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  3.58it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.84it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.04it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.15it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.25it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:02<00:01,  4.38it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.43it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.47it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.45it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.43it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.64it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.18it/s]
+Evaluated 400 examples
+[2026-07-22 17:30:32,675][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:13,  1.11s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:12,  1.11s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:11,  1.14s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:10,  1.14s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:09,  1.14s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.13s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.12s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.10s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:10<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:11<00:03,  1.10s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:12<00:02,  1.11s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.12s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.06s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.10s/it]
+Evaluated 400 examples
+[2026-07-22 17:30:47,039][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-22 17:30:47,039][evaluator][INFO] - Result for metric forget_truth_ratio:	0.6099912728131717
+[2026-07-22 17:30:47,044][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-22 17:30:47,044][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:30:47,051][metrics][INFO] - Evaluating forget_quality
+[2026-07-22 17:30:47,051][evaluator][INFO] - Result for metric forget_quality:	0.010032513978788836
+[2026-07-22 17:30:47,486][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.47it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.48it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.34it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:02,  4.35it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.39it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.44it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.44it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.54it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.58it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.58it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.55it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.52it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.75it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.55it/s]
+Evaluated 400 examples
+[2026-07-22 17:30:50,346][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.10463855743408203
+[2026-07-22 17:30:50,678][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:05,  2.32it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:04,  2.26it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:01<00:04,  2.28it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:04,  2.08it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:02<00:03,  2.30it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:02<00:02,  2.59it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:02<00:02,  2.81it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:03<00:01,  2.97it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:03<00:01,  3.03it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:03<00:01,  2.94it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:04<00:00,  3.02it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:04<00:00,  3.10it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:04<00:00,  3.05it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:04<00:00,  2.75it/s]
+Evaluated 400 examples
+[2026-07-22 17:30:55,404][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.2822923893033191
+[2026-07-22 17:30:55,726][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.68it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.74it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.75it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.77it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.77it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.70it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.69it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.66it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.69it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.68it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.62it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.65it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.87it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.74it/s]
+Evaluated 400 examples
+[2026-07-22 17:30:58,776][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:03,  3.01it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:03,  2.86it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:01<00:03,  2.81it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:02,  3.18it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:02,  2.67it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:02<00:02,  2.81it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:02<00:02,  2.63it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  2.55it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:03<00:01,  2.61it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:03<00:01,  2.77it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:04<00:00,  2.36it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:04<00:00,  2.47it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:04<00:00,  2.84it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:04<00:00,  2.71it/s]
+Evaluated 400 examples
+[2026-07-22 17:31:03,877][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.62it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.63it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.60it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.65it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.68it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.59it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.55it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.50it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.50it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.52it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.49it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.52it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.76it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.61it/s]
+Evaluated 400 examples
+[2026-07-22 17:31:07,003][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:12,  1.07s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:11,  1.07s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:10,  1.08s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:09,  1.07s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:08,  1.07s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.08s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.09s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.09s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:09<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:10<00:03,  1.10s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:11<00:02,  1.11s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.11s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.06s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.08s/it]
+Evaluated 400 examples
+[2026-07-22 17:31:21,045][metrics][INFO] - Evaluating retain_Truth_Ratio
+[2026-07-22 17:31:21,487][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.21it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.60it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  7.67it/s]
+Evaluated 100 examples
+[2026-07-22 17:31:22,810][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.37it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.50it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.08it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.46it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.51it/s]
+Evaluated 100 examples
+[2026-07-22 17:31:24,406][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+[2026-07-22 17:31:24,728][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  3.63it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  4.26it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  4.40it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  5.55it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  4.95it/s]
+Evaluated 100 examples
+[2026-07-22 17:31:25,537][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-22 17:31:25,537][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:31:25,537][metrics][INFO] - Evaluating ra_Truth_Ratio
+[2026-07-22 17:31:25,849][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  7.98it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.53it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  8.20it/s]
+Evaluated 117 examples
+[2026-07-22 17:31:26,653][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.51it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.48it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.06it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.51it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.53it/s]
+Evaluated 117 examples
+[2026-07-22 17:31:28,233][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+[2026-07-22 17:31:28,572][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  3.76it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  3.93it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  3.83it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  3.38it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  3.53it/s]
+Evaluated 117 examples
+[2026-07-22 17:31:29,704][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-22 17:31:29,704][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:31:29,705][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-22 17:31:29,705][metrics][INFO] - Evaluating model_utility
+[2026-07-22 17:31:29,705][evaluator][INFO] - Result for metric model_utility:	0.5407682812835304
+[2026-07-22 17:31:30,371][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:31:30,378][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.79it/s]
 31%|███       | 4/13 [00:00<00:00, 12.33it/s]
 46%|████▌     | 6/13 [00:00<00:00, 12.90it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.19it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.54it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.25it/s]
100%|██████████| 13/13 [00:00<00:00, 13.72it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 12.81it/s]
 31%|███       | 4/13 [00:00<00:00, 12.74it/s]
 46%|████▌     | 6/13 [00:00<00:00, 13.16it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.15it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.04it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.16it/s]
100%|██████████| 13/13 [00:00<00:00, 13.52it/s]
+[2026-07-22 17:31:32,290][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:31:32,298][metrics][INFO] - Evaluating privleak
+[2026-07-22 17:31:32,298][evaluator][INFO] - Result for metric privleak:	-3.3386649384157043
+[2026-07-22 17:31:32,629][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:00, 11.03it/s]
Calculating ES:  31%|███       | 4/13 [00:00<00:00, 10.09it/s]
Calculating ES:  46%|████▌     | 6/13 [00:00<00:00, 10.72it/s]
Calculating ES:  62%|██████▏   | 8/13 [00:00<00:00, 11.19it/s]
Calculating ES:  77%|███████▋  | 10/13 [00:00<00:00, 11.33it/s]
Calculating ES:  92%|█████████▏| 12/13 [00:01<00:00, 11.11it/s]
Calculating ES: 100%|██████████| 13/13 [00:01<00:00, 11.46it/s]
+Evaluated 400 examples
+[2026-07-22 17:31:33,764][evaluator][INFO] - Result for metric extraction_strength:	0.08332082050477542
diff --git a/logs/eval_NPO_lr2e-05_beta0.1_alpha2_epoch10.log b/logs/eval_NPO_lr2e-05_beta0.1_alpha2_epoch10.log
new file mode 100644
index 0000000000000000000000000000000000000000..0743f99a59b570d56013a8581b0e62600814c9f4
--- /dev/null
+++ b/logs/eval_NPO_lr2e-05_beta0.1_alpha2_epoch10.log
@@ -0,0 +1,78 @@
+/opt/conda/envs/copygen/lib/python3.12/site-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.6.3) or chardet (6.0.0.post1)/charset_normalizer (3.4.4) doesn't match a supported version!
+  warnings.warn(
+[2026-07-22 17:28:18,999][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/NPO_lr2e-05_beta0.1_alpha2_epoch10
+[2026-07-22 17:28:19,000][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-22 17:28:19,000][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/NPO_lr2e-05_beta0.1_alpha2_epoch10/TOFU_EVAL.json
+[2026-07-22 17:28:19,000][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/NPO_lr2e-05_beta0.1_alpha2_epoch10/TOFU_SUMMARY.json
+[2026-07-22 17:28:19,793][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:05,  2.18it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:03,  3.10it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  3.47it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.75it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:02,  3.96it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.07it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.18it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:02<00:01,  4.31it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.35it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.39it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.37it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.34it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.53it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.08it/s]
+Evaluated 400 examples
+[2026-07-22 17:28:23,719][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:13,  1.14s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:12,  1.13s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:11,  1.15s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:10,  1.15s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:09,  1.14s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.13s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.13s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:09<00:05,  1.11s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:10<00:04,  1.11s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:11<00:03,  1.11s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:12<00:02,  1.12s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.12s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.06s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.11s/it]
+Evaluated 400 examples
+[2026-07-22 17:28:38,138][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-22 17:28:38,138][evaluator][INFO] - Result for metric forget_truth_ratio:	0.6092674861552554
+[2026-07-22 17:28:38,143][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-22 17:28:38,143][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:28:38,150][metrics][INFO] - Evaluating forget_quality
+[2026-07-22 17:28:38,151][evaluator][INFO] - Result for metric forget_quality:	0.02424717195471133
+[2026-07-22 17:28:38,590][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.41it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.41it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.28it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:02,  4.31it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.35it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.40it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.37it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.45it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.50it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.50it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.48it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.41it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.62it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.46it/s]
+Evaluated 400 examples
+[2026-07-22 17:28:41,504][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.09674116134643555
+[2026-07-22 17:28:41,832][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:04,  2.62it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:03,  2.99it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:01<00:03,  2.86it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:03,  2.79it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:02,  2.87it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:02<00:02,  3.09it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:02<00:02,  2.98it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  3.14it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:03<00:01,  3.00it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:03<00:01,  2.75it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:03<00:00,  2.77it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:04<00:00,  2.87it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:04<00:00,  3.00it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:04<00:00,  2.93it/s]
+Evaluated 400 examples
+[2026-07-22 17:28:46,273][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.28020779717837746
+[2026-07-22 17:28:46,652][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.62it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.67it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.69it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.71it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.72it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.65it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.64it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.61it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.64it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.63it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.57it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.60it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.83it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.69it/s]
+Evaluated 400 examples
+[2026-07-22 17:28:49,737][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:03,  3.08it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:04,  2.66it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:01<00:03,  2.84it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:02,  3.23it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:02,  3.47it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:01<00:02,  3.26it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:02<00:01,  3.25it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  3.11it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:02<00:01,  3.06it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:03<00:00,  3.13it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:03<00:00,  3.00it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:03<00:00,  3.13it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:04<00:00,  3.39it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:04<00:00,  3.18it/s]
+Evaluated 400 examples
+[2026-07-22 17:28:54,164][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.61it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.61it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.57it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.63it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.66it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.58it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.54it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.48it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.47it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.49it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.45it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.48it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.72it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.58it/s]
+Evaluated 400 examples
+[2026-07-22 17:28:57,314][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:13,  1.09s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:11,  1.08s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:10,  1.08s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:09,  1.07s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:08,  1.06s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.08s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.09s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.10s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:09<00:04,  1.11s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:10<00:03,  1.10s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:12<00:02,  1.10s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.10s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.05s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.08s/it]
+Evaluated 400 examples
+[2026-07-22 17:29:11,358][metrics][INFO] - Evaluating retain_Truth_Ratio
+[2026-07-22 17:29:11,785][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.15it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.53it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  7.52it/s]
+Evaluated 100 examples
+[2026-07-22 17:29:12,768][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  5.83it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.46it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.06it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.08it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.22it/s]
+Evaluated 100 examples
+[2026-07-22 17:29:14,566][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+[2026-07-22 17:29:14,990][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  5.33it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  4.75it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  4.93it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  6.05it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  5.59it/s]
+Evaluated 100 examples
+[2026-07-22 17:29:15,706][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-22 17:29:15,707][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:29:15,707][metrics][INFO] - Evaluating ra_Truth_Ratio
+[2026-07-22 17:29:16,006][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.14it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.62it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  8.39it/s]
+Evaluated 117 examples
+[2026-07-22 17:29:16,866][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.78it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.50it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.10it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.57it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.59it/s]
+Evaluated 117 examples
+[2026-07-22 17:29:18,413][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+[2026-07-22 17:29:18,721][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  3.54it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  3.78it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  3.41it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  3.17it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  3.30it/s]
+Evaluated 117 examples
+[2026-07-22 17:29:19,933][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-22 17:29:19,933][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:29:19,933][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-22 17:29:19,933][metrics][INFO] - Evaluating model_utility
+[2026-07-22 17:29:19,933][evaluator][INFO] - Result for metric model_utility:	0.5814758976392439
+[2026-07-22 17:29:20,584][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:29:20,591][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.76it/s]
 31%|███       | 4/13 [00:00<00:00, 12.45it/s]
 46%|████▌     | 6/13 [00:00<00:00, 13.02it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.40it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.59it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.02it/s]
100%|██████████| 13/13 [00:00<00:00, 13.71it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 12.94it/s]
 31%|███       | 4/13 [00:00<00:00, 12.52it/s]
 46%|████▌     | 6/13 [00:00<00:00, 12.97it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.23it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.13it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.13it/s]
100%|██████████| 13/13 [00:00<00:00, 13.48it/s]
+[2026-07-22 17:29:22,506][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:29:22,514][metrics][INFO] - Evaluating privleak
+[2026-07-22 17:29:22,515][evaluator][INFO] - Result for metric privleak:	-10.865846002572797
+[2026-07-22 17:29:22,843][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:00, 11.45it/s]
Calculating ES:  31%|███       | 4/13 [00:00<00:00, 10.54it/s]
Calculating ES:  46%|████▌     | 6/13 [00:00<00:00, 11.25it/s]
Calculating ES:  62%|██████▏   | 8/13 [00:00<00:00, 11.72it/s]
Calculating ES:  77%|███████▋  | 10/13 [00:00<00:00, 11.93it/s]
Calculating ES:  92%|█████████▏| 12/13 [00:01<00:00, 11.75it/s]
Calculating ES: 100%|██████████| 13/13 [00:01<00:00, 12.08it/s]
+Evaluated 400 examples
+[2026-07-22 17:29:23,920][evaluator][INFO] - Result for metric extraction_strength:	0.08574521515981966
diff --git a/logs/eval_NPO_lr2e-05_beta0.1_alpha2_epoch5.log b/logs/eval_NPO_lr2e-05_beta0.1_alpha2_epoch5.log
new file mode 100644
index 0000000000000000000000000000000000000000..abc1357466902bce9ef5f1abe9e9267fe7c06156
--- /dev/null
+++ b/logs/eval_NPO_lr2e-05_beta0.1_alpha2_epoch5.log
@@ -0,0 +1,78 @@
+/opt/conda/envs/copygen/lib/python3.12/site-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.6.3) or chardet (6.0.0.post1)/charset_normalizer (3.4.4) doesn't match a supported version!
+  warnings.warn(
+[2026-07-22 17:36:25,367][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/NPO_lr2e-05_beta0.1_alpha2_epoch5
+[2026-07-22 17:36:25,368][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-22 17:36:25,368][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/NPO_lr2e-05_beta0.1_alpha2_epoch5/TOFU_EVAL.json
+[2026-07-22 17:36:25,368][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/NPO_lr2e-05_beta0.1_alpha2_epoch5/TOFU_SUMMARY.json
+[2026-07-22 17:36:26,183][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:05,  2.31it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:03,  3.21it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  3.59it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.83it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:02,  4.00it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.12it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.22it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:02<00:01,  4.33it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.37it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.39it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.38it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.36it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.56it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.14it/s]
+Evaluated 400 examples
+[2026-07-22 17:36:29,728][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:13,  1.12s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:12,  1.12s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:11,  1.14s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:10,  1.15s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:09,  1.13s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.13s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.13s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:09<00:05,  1.12s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:10<00:04,  1.11s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:11<00:03,  1.10s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:12<00:02,  1.11s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.12s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.06s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.11s/it]
+Evaluated 400 examples
+[2026-07-22 17:36:44,118][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-22 17:36:44,118][evaluator][INFO] - Result for metric forget_truth_ratio:	0.6223177600637857
+[2026-07-22 17:36:44,123][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-22 17:36:44,123][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:36:44,130][metrics][INFO] - Evaluating forget_quality
+[2026-07-22 17:36:44,131][evaluator][INFO] - Result for metric forget_quality:	0.054052912731737235
+[2026-07-22 17:36:44,548][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.46it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.46it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.30it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:02,  4.33it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.35it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.40it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.35it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.42it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.46it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.46it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.45it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.45it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.71it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.48it/s]
+Evaluated 400 examples
+[2026-07-22 17:36:47,450][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.10956892013549804
+[2026-07-22 17:36:47,760][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:04,  2.89it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:04,  2.55it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:01<00:03,  2.62it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:03,  2.78it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:02,  2.73it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:02<00:02,  2.98it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:02<00:01,  3.11it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  3.32it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:02<00:01,  3.29it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:03<00:00,  3.11it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:03<00:00,  3.04it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:04<00:00,  2.91it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:04<00:00,  3.04it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:04<00:00,  2.98it/s]
+Evaluated 400 examples
+[2026-07-22 17:36:52,117][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.28140295169525137
+[2026-07-22 17:36:52,448][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.58it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.66it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.69it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.72it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.73it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.67it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.65it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.62it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.61it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.56it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.49it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.53it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.77it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.66it/s]
+Evaluated 400 examples
+[2026-07-22 17:36:55,567][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:03,  3.60it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:02,  3.70it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:02,  3.57it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:02,  3.57it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:02,  3.70it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:01<00:01,  3.55it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:02<00:01,  3.36it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  3.20it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:02<00:01,  3.29it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:02<00:00,  3.20it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:03<00:00,  2.86it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:03<00:00,  2.99it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.12it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:03<00:00,  3.26it/s]
+Evaluated 400 examples
+[2026-07-22 17:36:59,861][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.59it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.60it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.57it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.62it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.64it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.56it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.53it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.48it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.47it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.49it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.47it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.49it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.68it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.57it/s]
+Evaluated 400 examples
+[2026-07-22 17:37:03,022][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:13,  1.09s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:11,  1.07s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:10,  1.08s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:09,  1.07s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:08,  1.06s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.08s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.09s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.10s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:09<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:10<00:03,  1.10s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:11<00:02,  1.10s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.10s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.05s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.08s/it]
+Evaluated 400 examples
+[2026-07-22 17:37:17,036][metrics][INFO] - Evaluating retain_Truth_Ratio
+[2026-07-22 17:37:17,451][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.04it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.48it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  7.46it/s]
+Evaluated 100 examples
+[2026-07-22 17:37:18,292][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.19it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.43it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.05it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.45it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.48it/s]
+Evaluated 100 examples
+[2026-07-22 17:37:19,905][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+[2026-07-22 17:37:20,229][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  6.14it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  5.37it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  5.08it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  6.27it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  5.91it/s]
+Evaluated 100 examples
+[2026-07-22 17:37:20,907][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-22 17:37:20,907][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:37:20,907][metrics][INFO] - Evaluating ra_Truth_Ratio
+[2026-07-22 17:37:21,240][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.05it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.58it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  8.35it/s]
+Evaluated 117 examples
+[2026-07-22 17:37:22,015][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.88it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.51it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.10it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.56it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.58it/s]
+Evaluated 117 examples
+[2026-07-22 17:37:23,564][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+[2026-07-22 17:37:23,867][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  3.67it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  3.91it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  3.68it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  3.81it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  3.79it/s]
+Evaluated 117 examples
+[2026-07-22 17:37:24,923][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-22 17:37:24,924][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:37:24,924][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-22 17:37:24,924][metrics][INFO] - Evaluating model_utility
+[2026-07-22 17:37:24,924][evaluator][INFO] - Result for metric model_utility:	0.5654685382532012
+[2026-07-22 17:37:25,562][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:37:25,570][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.91it/s]
 31%|███       | 4/13 [00:00<00:00, 12.61it/s]
 46%|████▌     | 6/13 [00:00<00:00, 13.16it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.26it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.67it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.60it/s]
100%|██████████| 13/13 [00:00<00:00, 13.89it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.01it/s]
 31%|███       | 4/13 [00:00<00:00, 12.73it/s]
 46%|████▌     | 6/13 [00:00<00:00, 13.06it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.20it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.42it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.52it/s]
100%|██████████| 13/13 [00:00<00:00, 13.76it/s]
+[2026-07-22 17:37:27,453][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:37:27,461][metrics][INFO] - Evaluating privleak
+[2026-07-22 17:37:27,461][evaluator][INFO] - Result for metric privleak:	-10.721588951716443
+[2026-07-22 17:37:27,801][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:00, 11.59it/s]
Calculating ES:  31%|███       | 4/13 [00:00<00:00, 10.44it/s]
Calculating ES:  46%|████▌     | 6/13 [00:00<00:00, 11.21it/s]
Calculating ES:  62%|██████▏   | 8/13 [00:00<00:00, 11.72it/s]
Calculating ES:  77%|███████▋  | 10/13 [00:00<00:00, 11.92it/s]
Calculating ES:  92%|█████████▏| 12/13 [00:01<00:00, 11.71it/s]
Calculating ES: 100%|██████████| 13/13 [00:01<00:00, 12.05it/s]
+Evaluated 400 examples
+[2026-07-22 17:37:28,881][evaluator][INFO] - Result for metric extraction_strength:	0.08470622527269665
diff --git a/logs/eval_NPO_lr2e-05_beta0.1_alpha5_epoch10.log b/logs/eval_NPO_lr2e-05_beta0.1_alpha5_epoch10.log
new file mode 100644
index 0000000000000000000000000000000000000000..291a47fdd22ef5a98e53a9a854af0e1d9597e0a4
--- /dev/null
+++ b/logs/eval_NPO_lr2e-05_beta0.1_alpha5_epoch10.log
@@ -0,0 +1,78 @@
+/opt/conda/envs/copygen/lib/python3.12/site-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.6.3) or chardet (6.0.0.post1)/charset_normalizer (3.4.4) doesn't match a supported version!
+  warnings.warn(
+[2026-07-22 17:32:04,245][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/NPO_lr2e-05_beta0.1_alpha5_epoch10
+[2026-07-22 17:32:04,246][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-22 17:32:04,246][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/NPO_lr2e-05_beta0.1_alpha5_epoch10/TOFU_EVAL.json
+[2026-07-22 17:32:04,246][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/NPO_lr2e-05_beta0.1_alpha5_epoch10/TOFU_SUMMARY.json
+[2026-07-22 17:32:05,039][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:04,  2.42it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:03,  3.29it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  3.62it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.86it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.06it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.16it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.26it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:02<00:01,  4.38it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.42it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.45it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.42it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.40it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.60it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.19it/s]
+Evaluated 400 examples
+[2026-07-22 17:32:08,554][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:13,  1.14s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:12,  1.13s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:11,  1.15s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:10,  1.15s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:09,  1.13s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.13s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.12s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.10s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:10<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:11<00:03,  1.10s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:12<00:02,  1.11s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.11s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.06s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.10s/it]
+Evaluated 400 examples
+[2026-07-22 17:32:22,907][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-22 17:32:22,908][evaluator][INFO] - Result for metric forget_truth_ratio:	0.6179996640683928
+[2026-07-22 17:32:22,912][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-22 17:32:22,912][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:32:22,919][metrics][INFO] - Evaluating forget_quality
+[2026-07-22 17:32:22,920][evaluator][INFO] - Result for metric forget_quality:	0.06521994354498474
+[2026-07-22 17:32:23,321][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.45it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.46it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.31it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:02,  4.34it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.38it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.43it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.43it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.49it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.52it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.51it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.49it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.43it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.63it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.49it/s]
+Evaluated 400 examples
+[2026-07-22 17:32:26,219][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.12286243438720704
+[2026-07-22 17:32:26,521][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:03,  3.12it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:04,  2.35it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:01<00:03,  2.58it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:03,  2.72it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:02,  2.80it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:02<00:02,  2.93it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:02<00:02,  2.86it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  2.97it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:03<00:01,  3.04it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:03<00:00,  3.06it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:03<00:00,  3.01it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:04<00:00,  2.85it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:04<00:00,  2.91it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:04<00:00,  2.87it/s]
+Evaluated 400 examples
+[2026-07-22 17:32:31,045][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.324568424555215
+[2026-07-22 17:32:31,378][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.64it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.69it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.70it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.72it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.73it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.66it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.64it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.61it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.64it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.64it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.57it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.61it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.84it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.70it/s]
+Evaluated 400 examples
+[2026-07-22 17:32:34,528][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:03,  3.03it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:03,  2.79it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:01<00:03,  2.87it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:02,  3.22it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:02,  3.29it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:01<00:02,  3.23it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:02<00:01,  3.17it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  3.15it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:02<00:01,  2.99it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:03<00:01,  2.96it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:03<00:00,  2.92it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:03<00:00,  3.10it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:04<00:00,  2.77it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:04<00:00,  2.99it/s]
+Evaluated 400 examples
+[2026-07-22 17:32:39,209][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.51it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.57it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.55it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.63it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.65it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.57it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.53it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.48it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.47it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.49it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.46it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.50it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.74it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.58it/s]
+Evaluated 400 examples
+[2026-07-22 17:32:42,387][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:13,  1.10s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:11,  1.08s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:10,  1.08s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:09,  1.07s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:08,  1.06s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.08s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.09s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.10s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:09<00:04,  1.11s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:10<00:03,  1.10s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:12<00:02,  1.11s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.10s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.05s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.08s/it]
+Evaluated 400 examples
+[2026-07-22 17:32:56,438][metrics][INFO] - Evaluating retain_Truth_Ratio
+[2026-07-22 17:32:57,202][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  7.99it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.46it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  7.46it/s]
+Evaluated 100 examples
+[2026-07-22 17:32:58,194][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.16it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.46it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.07it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.13it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.27it/s]
+Evaluated 100 examples
+[2026-07-22 17:32:59,956][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+[2026-07-22 17:33:00,265][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  4.24it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  4.63it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  4.89it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  5.81it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  5.33it/s]
+Evaluated 100 examples
+[2026-07-22 17:33:01,017][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-22 17:33:01,017][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:33:01,017][metrics][INFO] - Evaluating ra_Truth_Ratio
+[2026-07-22 17:33:01,320][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.13it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.62it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  8.39it/s]
+Evaluated 117 examples
+[2026-07-22 17:33:02,113][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.77it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.51it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.10it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.57it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.59it/s]
+Evaluated 117 examples
+[2026-07-22 17:33:03,659][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+[2026-07-22 17:33:04,060][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  3.53it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  3.74it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  3.21it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  3.44it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  3.44it/s]
+Evaluated 117 examples
+[2026-07-22 17:33:05,224][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-22 17:33:05,224][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:33:05,224][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-22 17:33:05,224][metrics][INFO] - Evaluating model_utility
+[2026-07-22 17:33:05,224][evaluator][INFO] - Result for metric model_utility:	0.5897317892611674
+[2026-07-22 17:33:05,880][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:33:05,892][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.01it/s]
 31%|███       | 4/13 [00:00<00:00, 11.82it/s]
 46%|████▌     | 6/13 [00:00<00:00, 12.43it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 12.89it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.26it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 12.93it/s]
100%|██████████| 13/13 [00:00<00:00, 13.38it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 12.87it/s]
 31%|███       | 4/13 [00:00<00:00, 12.53it/s]
 46%|████▌     | 6/13 [00:00<00:00, 12.69it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 12.91it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 12.78it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 12.90it/s]
100%|██████████| 13/13 [00:00<00:00, 13.24it/s]
+[2026-07-22 17:33:07,848][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:33:07,855][metrics][INFO] - Evaluating privleak
+[2026-07-22 17:33:07,855][evaluator][INFO] - Result for metric privleak:	-34.88489805270029
+[2026-07-22 17:33:08,191][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:01, 10.62it/s]
Calculating ES:  31%|███       | 4/13 [00:00<00:00,  9.95it/s]
Calculating ES:  46%|████▌     | 6/13 [00:00<00:00, 10.70it/s]
Calculating ES:  62%|██████▏   | 8/13 [00:00<00:00, 11.23it/s]
Calculating ES:  77%|███████▋  | 10/13 [00:00<00:00, 11.47it/s]
Calculating ES:  92%|█████████▏| 12/13 [00:01<00:00, 11.34it/s]
Calculating ES: 100%|██████████| 13/13 [00:01<00:00, 11.57it/s]
+Evaluated 400 examples
+[2026-07-22 17:33:09,315][evaluator][INFO] - Result for metric extraction_strength:	0.08082470260187125
diff --git a/logs/eval_NPO_lr2e-05_beta0.1_alpha5_epoch5.log b/logs/eval_NPO_lr2e-05_beta0.1_alpha5_epoch5.log
new file mode 100644
index 0000000000000000000000000000000000000000..eaf1845eea4a34ff52ef13a7636829c6db62b583
--- /dev/null
+++ b/logs/eval_NPO_lr2e-05_beta0.1_alpha5_epoch5.log
@@ -0,0 +1,78 @@
+/opt/conda/envs/copygen/lib/python3.12/site-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.6.3) or chardet (6.0.0.post1)/charset_normalizer (3.4.4) doesn't match a supported version!
+  warnings.warn(
+[2026-07-22 17:29:52,713][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/NPO_lr2e-05_beta0.1_alpha5_epoch5
+[2026-07-22 17:29:52,714][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-22 17:29:52,714][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/NPO_lr2e-05_beta0.1_alpha5_epoch5/TOFU_EVAL.json
+[2026-07-22 17:29:52,714][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/NPO_lr2e-05_beta0.1_alpha5_epoch5/TOFU_SUMMARY.json
+[2026-07-22 17:29:53,500][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:05,  2.24it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:03,  3.14it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  3.53it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.77it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:02,  3.98it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.10it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.21it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:02<00:01,  4.33it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.34it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.36it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.33it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.30it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.53it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.10it/s]
+Evaluated 400 examples
+[2026-07-22 17:29:57,131][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:13,  1.11s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:12,  1.11s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:11,  1.15s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:10,  1.15s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:09,  1.13s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.13s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.13s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.11s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:10<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:11<00:03,  1.10s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:12<00:02,  1.11s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.12s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.06s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.11s/it]
+Evaluated 400 examples
+[2026-07-22 17:30:11,522][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-22 17:30:11,523][evaluator][INFO] - Result for metric forget_truth_ratio:	0.620732403164572
+[2026-07-22 17:30:11,527][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-22 17:30:11,528][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:30:11,536][metrics][INFO] - Evaluating forget_quality
+[2026-07-22 17:30:11,537][evaluator][INFO] - Result for metric forget_quality:	0.036569856408414286
+[2026-07-22 17:30:11,952][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.46it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.38it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.22it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:02,  4.23it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.28it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.34it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.36it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.47it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.52it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.52it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.50it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.49it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.73it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.48it/s]
+Evaluated 400 examples
+[2026-07-22 17:30:14,853][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.13889686584472657
+[2026-07-22 17:30:15,194][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:04,  2.49it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:05,  2.19it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:01<00:04,  2.28it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:03,  2.49it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:02<00:03,  2.56it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:02<00:02,  2.62it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:02<00:02,  2.56it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:03<00:01,  2.66it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:03<00:01,  2.79it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:03<00:01,  2.68it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:04<00:00,  2.45it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:04<00:00,  2.47it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:05<00:00,  2.56it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:05<00:00,  2.54it/s]
+Evaluated 400 examples
+[2026-07-22 17:30:20,313][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.3109433894493367
+[2026-07-22 17:30:20,631][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.54it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.56it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.57it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.60it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.65it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.61it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.62it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.60it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.64it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.63it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.58it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.62it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.84it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.67it/s]
+Evaluated 400 examples
+[2026-07-22 17:30:23,728][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:04,  2.67it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:00<00:03,  3.22it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:00<00:03,  3.23it/s]
Calculating text similarity:  31%|███       | 4/13 [00:01<00:02,  3.32it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:01<00:02,  2.96it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:01<00:02,  3.05it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:02<00:02,  2.95it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:02<00:01,  3.10it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:02<00:01,  2.94it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:03<00:01,  2.85it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:03<00:00,  2.75it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:04<00:00,  2.89it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:04<00:00,  3.00it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:04<00:00,  2.99it/s]
+Evaluated 400 examples
+[2026-07-22 17:30:28,447][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.52it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.56it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.54it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.53it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.54it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.46it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.45it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.43it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.45it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.48it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.46it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.50it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.75it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.55it/s]
+Evaluated 400 examples
+[2026-07-22 17:30:31,612][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:12,  1.06s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:11,  1.06s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:10,  1.08s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:09,  1.07s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:08,  1.06s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.08s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.09s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.10s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:09<00:04,  1.10s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:10<00:03,  1.10s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:11<00:02,  1.11s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:13<00:01,  1.10s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.05s/it]
Calculating loss: 100%|██████████| 13/13 [00:14<00:00,  1.08s/it]
+Evaluated 400 examples
+[2026-07-22 17:30:45,624][metrics][INFO] - Evaluating retain_Truth_Ratio
+[2026-07-22 17:30:46,118][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  7.93it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.43it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  7.42it/s]
+Evaluated 100 examples
+[2026-07-22 17:30:46,959][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.46it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.50it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.09it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.50it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.54it/s]
+Evaluated 100 examples
+[2026-07-22 17:30:48,537][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+[2026-07-22 17:30:48,849][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  4.80it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  4.97it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  4.91it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  5.98it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:00<00:00,  5.54it/s]
+Evaluated 100 examples
+[2026-07-22 17:30:49,572][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-22 17:30:49,572][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:30:49,572][metrics][INFO] - Evaluating ra_Truth_Ratio
+[2026-07-22 17:30:49,872][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.12it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.60it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  8.32it/s]
+Evaluated 117 examples
+[2026-07-22 17:30:50,661][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.69it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.48it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.09it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.56it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.57it/s]
+Evaluated 117 examples
+[2026-07-22 17:30:52,217][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+[2026-07-22 17:30:52,655][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:00,  3.29it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  3.58it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:00<00:00,  3.23it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  3.41it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  3.39it/s]
+Evaluated 117 examples
+[2026-07-22 17:30:53,835][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-22 17:30:53,835][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-22 17:30:53,836][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-22 17:30:53,836][metrics][INFO] - Evaluating model_utility
+[2026-07-22 17:30:53,836][evaluator][INFO] - Result for metric model_utility:	0.5841298036164576
+[2026-07-22 17:30:54,536][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:30:54,548][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.49it/s]
 31%|███       | 4/13 [00:00<00:00, 12.36it/s]
 46%|████▌     | 6/13 [00:00<00:00, 13.08it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.47it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.81it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.54it/s]
100%|██████████| 13/13 [00:00<00:00, 13.97it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.19it/s]
 31%|███       | 4/13 [00:00<00:00, 13.15it/s]
 46%|████▌     | 6/13 [00:00<00:00, 13.62it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 13.68it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 13.75it/s]
 92%|█████████▏| 12/13 [00:00<00:00, 13.85it/s]
100%|██████████| 13/13 [00:00<00:00, 14.13it/s]
+[2026-07-22 17:30:56,401][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-22 17:30:56,408][metrics][INFO] - Evaluating privleak
+[2026-07-22 17:30:56,408][evaluator][INFO] - Result for metric privleak:	-37.35903301615938
+[2026-07-22 17:30:56,772][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:00, 11.78it/s]
Calculating ES:  31%|███       | 4/13 [00:00<00:00, 10.73it/s]
Calculating ES:  46%|████▌     | 6/13 [00:00<00:00, 11.38it/s]
Calculating ES:  62%|██████▏   | 8/13 [00:00<00:00, 11.80it/s]
Calculating ES:  77%|███████▋  | 10/13 [00:00<00:00, 11.94it/s]
Calculating ES:  92%|█████████▏| 12/13 [00:01<00:00, 11.75it/s]
Calculating ES: 100%|██████████| 13/13 [00:01<00:00, 12.14it/s]
+Evaluated 400 examples
+[2026-07-22 17:30:57,844][evaluator][INFO] - Result for metric extraction_strength:	0.08102853941546437
diff --git a/logs/eval_NPO_lr2e-05_beta0.5_alpha1_epoch10.log b/logs/eval_NPO_lr2e-05_beta0.5_alpha1_epoch10.log
new file mode 100644
index 0000000000000000000000000000000000000000..f1682200680b8b30df90729a14f529f5f3762dab
--- /dev/null
+++ b/logs/eval_NPO_lr2e-05_beta0.5_alpha1_epoch10.log
@@ -0,0 +1,78 @@
+/opt/conda/envs/copygen/lib/python3.12/site-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.6.3) or chardet (6.0.0.post1)/charset_normalizer (3.4.4) doesn't match a supported version!
+  warnings.warn(
+[2026-07-20 13:32:05,069][evaluator][INFO] - Evaluations stored in the experiment directory: ./saves/eval/NPO_lr2e-05_beta0.5_alpha1_epoch10
+[2026-07-20 13:32:05,070][evaluator][INFO] - ***** Running TOFU evaluation suite *****
+[2026-07-20 13:32:05,070][evaluator][INFO] - Fine-grained evaluations will be saved to: ./saves/eval/NPO_lr2e-05_beta0.5_alpha1_epoch10/TOFU_EVAL.json
+[2026-07-20 13:32:05,070][evaluator][INFO] - Aggregated evaluations will be summarised in: ./saves/eval/NPO_lr2e-05_beta0.5_alpha1_epoch10/TOFU_SUMMARY.json
+[2026-07-20 13:32:05,884][metrics][INFO] - Evaluating forget_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:08,  1.47it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:04,  2.48it/s]
Calculating loss:  23%|██▎       | 3/13 [00:01<00:03,  3.12it/s]
Calculating loss:  31%|███       | 4/13 [00:01<00:02,  3.59it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:02,  3.93it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.17it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.35it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:02<00:01,  4.53it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:02<00:00,  4.61it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.67it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.69it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.70it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.84it/s]
Calculating loss: 100%|██████████| 13/13 [00:03<00:00,  4.07it/s]
+Evaluated 400 examples
+[2026-07-20 13:32:09,503][metrics][INFO] - Evaluating forget_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:12,  1.07s/it]
Calculating loss:  15%|█▌        | 2/13 [00:02<00:11,  1.04s/it]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:10,  1.05s/it]
Calculating loss:  31%|███       | 4/13 [00:04<00:09,  1.05s/it]
Calculating loss:  38%|███▊      | 5/13 [00:05<00:08,  1.05s/it]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.04s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.03s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.02s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:09<00:04,  1.01s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:10<00:03,  1.01s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:11<00:02,  1.02s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:12<00:01,  1.02s/it]
Calculating loss: 100%|██████████| 13/13 [00:13<00:00,  1.02it/s]
Calculating loss: 100%|██████████| 13/13 [00:13<00:00,  1.02s/it]
+Evaluated 400 examples
+[2026-07-20 13:32:22,729][metrics][INFO] - Evaluating forget_truth_ratio
+[2026-07-20 13:32:22,730][evaluator][INFO] - Result for metric forget_truth_ratio:	0.5368269645950674
+[2026-07-20 13:32:22,738][metrics][INFO] - Skipping forget_quality's precompute forget_truth_ratio, already evaluated.
+[2026-07-20 13:32:22,739][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-20 13:32:22,757][metrics][INFO] - Evaluating forget_quality
+[2026-07-20 13:32:22,759][evaluator][INFO] - Result for metric forget_quality:	1.6196210141854015e-10
+[2026-07-20 13:32:23,131][metrics][INFO] - Evaluating forget_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.76it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.58it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.53it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.62it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.67it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.61it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.65it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.76it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.77it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.78it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.79it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.79it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  5.06it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.79it/s]
+Evaluated 400 examples
+[2026-07-20 13:32:25,849][evaluator][INFO] - Result for metric forget_Q_A_Prob:	0.40543426513671876
+[2026-07-20 13:32:26,133][metrics][INFO] - Evaluating forget_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:10,  1.17it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:01<00:08,  1.27it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:02<00:08,  1.22it/s]
Calculating text similarity:  31%|███       | 4/13 [00:03<00:06,  1.30it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:03<00:05,  1.38it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:04<00:04,  1.48it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:05<00:04,  1.38it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:05<00:03,  1.46it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:06<00:02,  1.57it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:07<00:02,  1.40it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:07<00:01,  1.43it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:08<00:00,  1.45it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:09<00:00,  1.47it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:09<00:00,  1.41it/s]
+Evaluated 400 examples
+[2026-07-20 13:32:35,342][evaluator][INFO] - Result for metric forget_Q_A_ROUGE:	0.42873284272535533
+[2026-07-20 13:32:35,718][metrics][INFO] - Evaluating retain_Q_A_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:03,  3.94it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.53it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.74it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.86it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  4.92it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.83it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.86it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.86it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.90it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.91it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.87it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.92it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.50it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.73it/s]
+Evaluated 400 examples
+[2026-07-20 13:32:38,724][metrics][INFO] - Evaluating retain_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating text similarity:   8%|▊         | 1/13 [00:00<00:06,  1.73it/s]
Calculating text similarity:  15%|█▌        | 2/13 [00:01<00:05,  1.83it/s]
Calculating text similarity:  23%|██▎       | 3/13 [00:01<00:05,  1.74it/s]
Calculating text similarity:  31%|███       | 4/13 [00:02<00:04,  1.85it/s]
Calculating text similarity:  38%|███▊      | 5/13 [00:02<00:04,  1.88it/s]
Calculating text similarity:  46%|████▌     | 6/13 [00:03<00:04,  1.74it/s]
Calculating text similarity:  54%|█████▍    | 7/13 [00:04<00:03,  1.59it/s]
Calculating text similarity:  62%|██████▏   | 8/13 [00:04<00:03,  1.58it/s]
Calculating text similarity:  69%|██████▉   | 9/13 [00:05<00:02,  1.63it/s]
Calculating text similarity:  77%|███████▋  | 10/13 [00:05<00:01,  1.62it/s]
Calculating text similarity:  85%|████████▍ | 11/13 [00:06<00:01,  1.51it/s]
Calculating text similarity:  92%|█████████▏| 12/13 [00:07<00:00,  1.55it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:07<00:00,  1.69it/s]
Calculating text similarity: 100%|██████████| 13/13 [00:07<00:00,  1.67it/s]
+Evaluated 400 examples
+[2026-07-20 13:32:47,025][metrics][INFO] - Evaluating retain_Q_A_PARA_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:00<00:02,  4.90it/s]
Calculating loss:  15%|█▌        | 2/13 [00:00<00:02,  4.93it/s]
Calculating loss:  23%|██▎       | 3/13 [00:00<00:02,  4.91it/s]
Calculating loss:  31%|███       | 4/13 [00:00<00:01,  4.96it/s]
Calculating loss:  38%|███▊      | 5/13 [00:01<00:01,  5.00it/s]
Calculating loss:  46%|████▌     | 6/13 [00:01<00:01,  4.92it/s]
Calculating loss:  54%|█████▍    | 7/13 [00:01<00:01,  4.89it/s]
Calculating loss:  62%|██████▏   | 8/13 [00:01<00:01,  4.84it/s]
Calculating loss:  69%|██████▉   | 9/13 [00:01<00:00,  4.83it/s]
Calculating loss:  77%|███████▋  | 10/13 [00:02<00:00,  4.76it/s]
Calculating loss:  85%|████████▍ | 11/13 [00:02<00:00,  4.76it/s]
Calculating loss:  92%|█████████▏| 12/13 [00:02<00:00,  4.79it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  5.07it/s]
Calculating loss: 100%|██████████| 13/13 [00:02<00:00,  4.91it/s]
+Evaluated 400 examples
+[2026-07-20 13:32:49,955][metrics][INFO] - Evaluating retain_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating loss:   8%|▊         | 1/13 [00:01<00:12,  1.02s/it]
Calculating loss:  15%|█▌        | 2/13 [00:01<00:10,  1.01it/s]
Calculating loss:  23%|██▎       | 3/13 [00:03<00:09,  1.00it/s]
Calculating loss:  31%|███       | 4/13 [00:03<00:08,  1.01it/s]
Calculating loss:  38%|███▊      | 5/13 [00:04<00:07,  1.01it/s]
Calculating loss:  46%|████▌     | 6/13 [00:06<00:07,  1.01s/it]
Calculating loss:  54%|█████▍    | 7/13 [00:07<00:06,  1.01s/it]
Calculating loss:  62%|██████▏   | 8/13 [00:08<00:05,  1.01s/it]
Calculating loss:  69%|██████▉   | 9/13 [00:09<00:04,  1.01s/it]
Calculating loss:  77%|███████▋  | 10/13 [00:10<00:03,  1.01s/it]
Calculating loss:  85%|████████▍ | 11/13 [00:11<00:02,  1.01s/it]
Calculating loss:  92%|█████████▏| 12/13 [00:12<00:01,  1.01s/it]
Calculating loss: 100%|██████████| 13/13 [00:12<00:00,  1.04it/s]
Calculating loss: 100%|██████████| 13/13 [00:12<00:00,  1.00it/s]
+Evaluated 400 examples
+[2026-07-20 13:33:02,902][metrics][INFO] - Evaluating retain_Truth_Ratio
+[2026-07-20 13:33:03,275][metrics][INFO] - Evaluating ra_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  7.54it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.48it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  7.57it/s]
+Evaluated 100 examples
+[2026-07-20 13:33:04,064][metrics][INFO] - Evaluating ra_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.04it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.60it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.14it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.56it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.60it/s]
+Evaluated 100 examples
+[2026-07-20 13:33:05,606][metrics][INFO] - Evaluating ra_Q_A_Prob_normalised
+[2026-07-20 13:33:05,877][metrics][INFO] - Evaluating ra_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:01,  2.92it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:00<00:00,  2.87it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:01<00:00,  2.91it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  3.32it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  3.15it/s]
+Evaluated 100 examples
+[2026-07-20 13:33:07,148][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_Prob, already evaluated.
+[2026-07-20 13:33:07,148][metrics][INFO] - Skipping ra_Truth_Ratio's precompute ra_Q_A_PERT_Prob, already evaluated.
+[2026-07-20 13:33:07,148][metrics][INFO] - Evaluating ra_Truth_Ratio
+[2026-07-20 13:33:07,410][metrics][INFO] - Evaluating wf_Q_A_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  8.31it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:00<00:00,  6.86it/s]
Calculating loss: 100%|██████████| 4/4 [00:00<00:00,  8.62it/s]
+Evaluated 117 examples
+[2026-07-20 13:33:08,268][metrics][INFO] - Evaluating wf_Q_A_PERT_Prob
+
Calculating loss:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating loss:  25%|██▌       | 1/4 [00:00<00:00,  6.38it/s]
Calculating loss:  50%|█████     | 2/4 [00:00<00:00,  2.61it/s]
Calculating loss:  75%|███████▌  | 3/4 [00:01<00:00,  2.22it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.70it/s]
Calculating loss: 100%|██████████| 4/4 [00:01<00:00,  2.71it/s]
+Evaluated 117 examples
+[2026-07-20 13:33:09,747][metrics][INFO] - Evaluating wf_Q_A_Prob_normalised
+[2026-07-20 13:33:10,032][metrics][INFO] - Evaluating wf_Q_A_ROUGE
+
Calculating text similarity:   0%|          | 0/4 [00:00<?, ?it/s]
Calculating text similarity:  25%|██▌       | 1/4 [00:00<00:01,  1.87it/s]
Calculating text similarity:  50%|█████     | 2/4 [00:01<00:00,  2.01it/s]
Calculating text similarity:  75%|███████▌  | 3/4 [00:01<00:00,  2.13it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  2.21it/s]
Calculating text similarity: 100%|██████████| 4/4 [00:01<00:00,  2.14it/s]
+Evaluated 117 examples
+[2026-07-20 13:33:11,902][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_Prob, already evaluated.
+[2026-07-20 13:33:11,902][metrics][INFO] - Skipping wf_Truth_Ratio's precompute wf_Q_A_PERT_Prob, already evaluated.
+[2026-07-20 13:33:11,902][metrics][INFO] - Evaluating wf_Truth_Ratio
+[2026-07-20 13:33:11,902][metrics][INFO] - Evaluating model_utility
+[2026-07-20 13:33:11,903][evaluator][INFO] - Result for metric model_utility:	0.5944411856791775
+[2026-07-20 13:33:12,489][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-20 13:33:12,504][metrics][INFO] - Evaluating mia_min_k
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 13.28it/s]
 31%|███       | 4/13 [00:00<00:00, 11.36it/s]
 46%|████▌     | 6/13 [00:00<00:00, 11.68it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 11.78it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 12.09it/s]
 92%|█████████▏| 12/13 [00:01<00:00, 11.78it/s]
100%|██████████| 13/13 [00:01<00:00, 12.33it/s]
+
  0%|          | 0/13 [00:00<?, ?it/s]
 15%|█▌        | 2/13 [00:00<00:00, 11.59it/s]
 31%|███       | 4/13 [00:00<00:00, 11.44it/s]
 46%|████▌     | 6/13 [00:00<00:00, 11.77it/s]
 62%|██████▏   | 8/13 [00:00<00:00, 11.00it/s]
 77%|███████▋  | 10/13 [00:00<00:00, 11.43it/s]
 92%|█████████▏| 12/13 [00:01<00:00, 11.72it/s]
100%|██████████| 13/13 [00:01<00:00, 11.97it/s]
+[2026-07-20 13:33:14,650][metrics][INFO] - Loading evaluations from /workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
+[2026-07-20 13:33:14,665][metrics][INFO] - Evaluating privleak
+[2026-07-20 13:33:14,665][evaluator][INFO] - Result for metric privleak:	-90.55192240034788
+[2026-07-20 13:33:14,965][metrics][INFO] - Evaluating extraction_strength
+
Calculating ES:   0%|          | 0/13 [00:00<?, ?it/s]
Calculating ES:  15%|█▌        | 2/13 [00:00<00:01, 10.25it/s]
Calculating ES:  31%|███       | 4/13 [00:00<00:00,  9.35it/s]
Calculating ES:  46%|████▌     | 6/13 [00:00<00:00,  9.81it/s]
Calculating ES:  54%|█████▍    | 7/13 [00:00<00:00,  9.79it/s]
Calculating ES:  69%|██████▉   | 9/13 [00:00<00:00, 10.28it/s]
Calculating ES:  85%|████████▍ | 11/13 [00:01<00:00, 10.10it/s]
Calculating ES: 100%|██████████| 13/13 [00:01<00:00, 10.84it/s]
Calculating ES: 100%|██████████| 13/13 [00:01<00:00, 10.31it/s]
+Evaluated 400 examples
+[2026-07-20 13:33:16,226][evaluator][INFO] - Result for metric extraction_strength:	0.13835912825501281
diff --git a/logs/fit_full.log b/logs/fit_full.log
new file mode 100644
index 0000000000000000000000000000000000000000..04b2fa11d094f2cf01ad06b30f3bf7666b1e6f3a
--- /dev/null
+++ b/logs/fit_full.log
@@ -0,0 +1,1022 @@
+
Loading weights:   0%|          | 0/146 [00:00<?, ?it/s]
Loading weights: 100%|██████████| 146/146 [00:00<00:00, 18496.64it/s]
+2026-07-20 12:54:04,348 HTTP Request: HEAD https://huggingface.co/datasets/Salesforce/wikitext/resolve/main/README.md "HTTP/1.1 307 Temporary Redirect"
+2026-07-20 12:54:04,359 HTTP Request: HEAD https://huggingface.co/api/resolve-cache/datasets/Salesforce/wikitext/b08601e04326c79dfdd32d625aee71d232d685c3/README.md "HTTP/1.1 200 OK"
+2026-07-20 12:54:04,399 HTTP Request: HEAD https://huggingface.co/datasets/Salesforce/wikitext/resolve/b08601e04326c79dfdd32d625aee71d232d685c3/wikitext.py "HTTP/1.1 404 Not Found"
+Warning: You are sending unauthenticated requests to the HF Hub. Please set a HF_TOKEN to enable higher rate limits and faster downloads.
+2026-07-20 12:54:04,399 Warning: You are sending unauthenticated requests to the HF Hub. Please set a HF_TOKEN to enable higher rate limits and faster downloads.
+2026-07-20 12:54:04,487 HTTP Request: HEAD https://s3.amazonaws.com/datasets.huggingface.co/datasets/datasets/Salesforce/wikitext/Salesforce/wikitext.py "HTTP/1.1 404 Not Found"
+2026-07-20 12:54:04,667 HTTP Request: GET https://huggingface.co/api/datasets/Salesforce/wikitext/revision/b08601e04326c79dfdd32d625aee71d232d685c3 "HTTP/1.1 200 OK"
+2026-07-20 12:54:04,763 HTTP Request: HEAD https://huggingface.co/datasets/Salesforce/wikitext/resolve/b08601e04326c79dfdd32d625aee71d232d685c3/.huggingface.yaml "HTTP/1.1 404 Not Found"
+2026-07-20 12:54:04,848 HTTP Request: GET https://datasets-server.huggingface.co/info?dataset=Salesforce/wikitext "HTTP/1.1 200 OK"
+2026-07-20 12:54:04,905 HTTP Request: GET https://huggingface.co/api/datasets/Salesforce/wikitext/tree/b08601e04326c79dfdd32d625aee71d232d685c3/wikitext-103-raw-v1?recursive=true&expand=false "HTTP/1.1 200 OK"
+2026-07-20 12:54:04,946 HTTP Request: GET https://huggingface.co/api/datasets/Salesforce/wikitext/tree/b08601e04326c79dfdd32d625aee71d232d685c3?recursive=false&expand=false "HTTP/1.1 200 OK"
+2026-07-20 12:54:04,993 HTTP Request: HEAD https://huggingface.co/datasets/Salesforce/wikitext/resolve/b08601e04326c79dfdd32d625aee71d232d685c3/dataset_infos.json "HTTP/1.1 404 Not Found"
+[transformers] Ignoring clean_up_tokenization_spaces=True for BPE tokenizer TokenizersBackend. The clean_up_tokenization post-processing step is designed for WordPiece tokenizers and is destructive for BPE (it strips spaces before punctuation). Set clean_up_tokenization_spaces=False to suppress this warning, or set clean_up_tokenization_spaces_for_bpe_even_though_it_will_corrupt_output=True to force cleanup anyway.
+2026-07-20 12:54:05,335 fit: n_layers=16 d_model=2048, fitting 15 source layers (target=L15) on 1000 prompts
+/opt/conda/envs/copygen/lib/python3.12/site-packages/torch/autograd/graph.py:829: UserWarning: Attempting to run cuBLAS, but there was no current CUDA context! Attempting to set the primary context... (Triggered internally at /pytorch/aten/src/ATen/cuda/CublasHandlePool.cpp:179.)
+  return Variable._execution_engine.run_backward(  # Calls into the C++ engine to run the backward pass
+2026-07-20 12:54:10,144   prompt 1/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.918  max_d_mean=nan
+2026-07-20 12:54:14,489   prompt 2/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=1.832  max_d_mean=5.81e-01
+2026-07-20 12:54:18,759   prompt 3/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=1.723  max_d_mean=4.00e-01
+2026-07-20 12:54:23,047   prompt 4/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=2.232  max_d_mean=3.85e-01
+2026-07-20 12:54:27,301   prompt 5/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=1.928  max_d_mean=2.84e-01
+2026-07-20 12:54:31,545   prompt 6/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=2.456  max_d_mean=3.10e-01
+2026-07-20 12:54:35,930   prompt 7/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=1.914  max_d_mean=2.11e-01
+2026-07-20 12:54:40,394   prompt 8/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=1.734  max_d_mean=1.73e-01
+2026-07-20 12:54:44,702   prompt 9/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=2.083  max_d_mean=1.93e-01
+2026-07-20 12:54:49,307   prompt 10/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.249  max_d_mean=1.94e-01
+2026-07-20 12:54:53,692   prompt 11/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=1.916  max_d_mean=1.52e-01
+2026-07-20 12:54:58,218   prompt 12/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.665  max_d_mean=1.22e-01
+2026-07-20 12:55:02,607   prompt 13/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=2.078  max_d_mean=1.38e-01
+2026-07-20 12:55:07,204   prompt 14/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.227  max_d_mean=1.35e-01
+2026-07-20 12:55:11,667   prompt 15/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=2.445  max_d_mean=1.40e-01
+2026-07-20 12:55:16,069   prompt 16/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=1.779  max_d_mean=9.17e-02
+2026-07-20 12:55:20,494   prompt 17/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=1.865  max_d_mean=9.73e-02
+2026-07-20 12:55:24,994   prompt 18/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=1.672  max_d_mean=6.75e-02
+2026-07-20 12:55:29,473   prompt 19/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=2.156  max_d_mean=9.30e-02
+2026-07-20 12:55:33,933   prompt 20/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=1.983  max_d_mean=9.03e-02
+2026-07-20 12:55:38,619   prompt 21/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.518  max_d_mean=6.79e-02
+2026-07-20 12:55:43,177   prompt 22/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.808  max_d_mean=6.95e-02
+2026-07-20 12:55:47,620   prompt 23/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=1.697  max_d_mean=6.20e-02
+2026-07-20 12:55:52,057   prompt 24/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=1.692  max_d_mean=6.05e-02
+2026-07-20 12:55:56,576   prompt 25/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.003  max_d_mean=6.59e-02
+2026-07-20 12:56:01,113   prompt 26/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=1.734  max_d_mean=5.74e-02
+2026-07-20 12:56:05,629   prompt 27/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.826  max_d_mean=5.47e-02
+2026-07-20 12:56:10,214   prompt 28/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.679  max_d_mean=5.15e-02
+2026-07-20 12:56:14,735   prompt 29/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.267  max_d_mean=6.88e-02
+2026-07-20 12:56:19,261   prompt 30/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.910  max_d_mean=5.89e-02
+2026-07-20 12:56:23,824   prompt 31/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.443  max_d_mean=3.90e-02
+2026-07-20 12:56:28,441   prompt 32/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.146  max_d_mean=6.35e-02
+2026-07-20 12:56:32,968   prompt 33/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.695  max_d_mean=4.74e-02
+2026-07-20 12:56:37,607   prompt 34/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.983  max_d_mean=5.72e-02
+2026-07-20 12:56:42,167   prompt 35/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.178  max_d_mean=6.20e-02
+2026-07-20 12:56:46,598   prompt 36/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=1.509  max_d_mean=3.56e-02
+2026-07-20 12:56:51,150   prompt 37/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.542  max_d_mean=3.95e-02
+2026-07-20 12:56:55,771   prompt 38/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.215  max_d_mean=5.71e-02
+2026-07-20 12:57:00,414   prompt 39/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.826  max_d_mean=4.61e-02
+2026-07-20 12:57:04,943   prompt 40/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.607  max_d_mean=3.85e-02
+2026-07-20 12:57:09,708   prompt 41/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.595  max_d_mean=3.73e-02
+2026-07-20 12:57:14,319   prompt 42/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.209  max_d_mean=5.34e-02
+2026-07-20 12:57:18,847   prompt 43/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.933  max_d_mean=4.47e-02
+2026-07-20 12:57:23,481   prompt 44/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.284  max_d_mean=5.38e-02
+2026-07-20 12:57:28,102   prompt 45/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.129  max_d_mean=4.89e-02
+2026-07-20 12:57:32,677   prompt 46/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.959  max_d_mean=4.22e-02
+2026-07-20 12:57:37,204   prompt 47/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.812  max_d_mean=3.91e-02
+2026-07-20 12:57:41,997   prompt 48/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=3.064  max_d_mean=6.84e-02
+2026-07-20 12:57:46,603   prompt 49/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.940  max_d_mean=6.15e-02
+2026-07-20 12:57:51,165   prompt 50/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.858  max_d_mean=3.93e-02
+2026-07-20 12:57:56,215   prompt 51/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.779  max_d_mean=3.43e-02
+2026-07-20 12:58:01,000   prompt 52/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.077  max_d_mean=4.06e-02
+2026-07-20 12:58:05,769   prompt 53/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.846  max_d_mean=3.58e-02
+2026-07-20 12:58:10,425   prompt 54/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.409  max_d_mean=4.62e-02
+2026-07-20 12:58:15,026   prompt 55/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.068  max_d_mean=3.87e-02
+2026-07-20 12:58:19,600   prompt 56/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.938  max_d_mean=3.65e-02
+2026-07-20 12:58:24,138   prompt 57/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.093  max_d_mean=3.66e-02
+2026-07-20 12:58:28,811   prompt 58/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.625  max_d_mean=2.92e-02
+2026-07-20 12:58:33,397   prompt 59/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.152  max_d_mean=3.72e-02
+2026-07-20 12:58:37,982   prompt 60/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=3.552  max_d_mean=6.26e-02
+2026-07-20 12:58:42,780   prompt 61/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.529  max_d_mean=4.18e-02
+2026-07-20 12:58:47,379   prompt 62/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.327  max_d_mean=3.62e-02
+2026-07-20 12:58:52,270   prompt 63/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.065  max_d_mean=3.11e-02
+2026-07-20 12:58:56,896   prompt 64/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.642  max_d_mean=2.45e-02
+2026-07-20 12:59:01,564   prompt 65/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.986  max_d_mean=3.07e-02
+2026-07-20 12:59:06,196   prompt 66/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.973  max_d_mean=2.97e-02
+2026-07-20 12:59:10,690   prompt 67/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=1.923  max_d_mean=2.77e-02
+2026-07-20 12:59:15,270   prompt 68/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.736  max_d_mean=2.53e-02
+2026-07-20 12:59:19,768   prompt 69/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=1.604  max_d_mean=2.12e-02
+2026-07-20 12:59:24,273   prompt 70/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.747  max_d_mean=2.39e-02
+2026-07-20 12:59:28,905   prompt 71/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.578  max_d_mean=2.03e-02
+2026-07-20 12:59:33,405   prompt 72/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=1.571  max_d_mean=1.97e-02
+2026-07-20 12:59:38,102   prompt 73/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.819  max_d_mean=2.54e-02
+2026-07-20 12:59:42,620   prompt 74/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.659  max_d_mean=2.28e-02
+2026-07-20 12:59:47,291   prompt 75/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.686  max_d_mean=2.30e-02
+2026-07-20 12:59:52,101   prompt 76/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.608  max_d_mean=2.10e-02
+2026-07-20 12:59:56,711   prompt 77/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.084  max_d_mean=2.74e-02
+2026-07-20 13:00:01,406   prompt 78/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.927  max_d_mean=2.53e-02
+2026-07-20 13:00:06,066   prompt 79/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.224  max_d_mean=2.78e-02
+2026-07-20 13:00:10,590   prompt 80/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.055  max_d_mean=2.59e-02
+2026-07-20 13:00:15,293   prompt 81/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.847  max_d_mean=2.36e-02
+2026-07-20 13:00:19,778   prompt 82/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=1.926  max_d_mean=2.31e-02
+2026-07-20 13:00:24,291   prompt 83/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.319  max_d_mean=2.80e-02
+2026-07-20 13:00:28,892   prompt 84/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.742  max_d_mean=2.15e-02
+2026-07-20 13:00:33,418   prompt 85/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.667  max_d_mean=1.92e-02
+2026-07-20 13:00:37,984   prompt 86/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.161  max_d_mean=2.58e-02
+2026-07-20 13:00:42,570   prompt 87/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.133  max_d_mean=2.51e-02
+2026-07-20 13:00:47,175   prompt 88/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.003  max_d_mean=2.30e-02
+2026-07-20 13:00:51,670   prompt 89/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=2.184  max_d_mean=2.55e-02
+2026-07-20 13:00:56,402   prompt 90/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.960  max_d_mean=2.19e-02
+2026-07-20 13:01:01,284   prompt 91/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.032  max_d_mean=2.28e-02
+2026-07-20 13:01:06,384   prompt 92/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.790  max_d_mean=2.03e-02
+2026-07-20 13:01:11,203   prompt 93/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.996  max_d_mean=2.12e-02
+2026-07-20 13:01:15,692   prompt 94/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=1.513  max_d_mean=1.75e-02
+2026-07-20 13:01:20,246   prompt 95/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.198  max_d_mean=2.45e-02
+2026-07-20 13:01:24,866   prompt 96/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.762  max_d_mean=1.94e-02
+2026-07-20 13:01:29,478   prompt 97/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.979  max_d_mean=2.24e-02
+2026-07-20 13:01:34,175   prompt 98/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.700  max_d_mean=1.86e-02
+2026-07-20 13:01:38,746   prompt 99/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.732  max_d_mean=1.91e-02
+2026-07-20 13:01:43,479   prompt 100/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.419  max_d_mean=2.60e-02
+2026-07-20 13:01:48,387   prompt 101/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.789  max_d_mean=1.94e-02
+2026-07-20 13:01:53,000   prompt 102/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.021  max_d_mean=2.09e-02
+2026-07-20 13:01:57,514   prompt 103/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.188  max_d_mean=2.26e-02
+2026-07-20 13:02:02,084   prompt 104/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.194  max_d_mean=2.25e-02
+2026-07-20 13:02:06,638   prompt 105/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.837  max_d_mean=1.81e-02
+2026-07-20 13:02:11,276   prompt 106/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.894  max_d_mean=1.95e-02
+2026-07-20 13:02:15,972   prompt 107/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.697  max_d_mean=1.52e-02
+2026-07-20 13:02:20,687   prompt 108/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.587  max_d_mean=1.19e-02
+2026-07-20 13:02:25,200   prompt 109/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.251  max_d_mean=2.19e-02
+2026-07-20 13:02:29,927   prompt 110/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.834  max_d_mean=1.79e-02
+2026-07-20 13:02:34,609   prompt 111/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.807  max_d_mean=1.77e-02
+2026-07-20 13:02:39,296   prompt 112/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.912  max_d_mean=1.82e-02
+2026-07-20 13:02:43,897   prompt 113/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.775  max_d_mean=1.70e-02
+2026-07-20 13:02:48,517   prompt 114/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.916  max_d_mean=2.72e-02
+2026-07-20 13:02:53,093   prompt 115/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.003  max_d_mean=1.92e-02
+2026-07-20 13:02:57,661   prompt 116/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.444  max_d_mean=2.26e-02
+2026-07-20 13:03:02,216   prompt 117/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.954  max_d_mean=1.76e-02
+2026-07-20 13:03:06,919   prompt 118/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.058  max_d_mean=1.80e-02
+2026-07-20 13:03:11,597   prompt 119/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.404  max_d_mean=2.12e-02
+2026-07-20 13:03:16,213   prompt 120/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.066  max_d_mean=1.80e-02
+2026-07-20 13:03:20,923   prompt 121/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.503  max_d_mean=1.30e-02
+2026-07-20 13:03:25,511   prompt 122/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.656  max_d_mean=1.19e-02
+2026-07-20 13:03:30,097   prompt 123/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.949  max_d_mean=1.69e-02
+2026-07-20 13:03:34,615   prompt 124/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.838  max_d_mean=1.51e-02
+2026-07-20 13:03:39,142   prompt 125/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.992  max_d_mean=1.74e-02
+2026-07-20 13:03:43,971   prompt 126/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.957  max_d_mean=1.65e-02
+2026-07-20 13:03:48,474   prompt 127/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.102  max_d_mean=1.75e-02
+2026-07-20 13:03:53,049   prompt 128/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.768  max_d_mean=2.27e-02
+2026-07-20 13:03:57,700   prompt 129/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.740  max_d_mean=1.43e-02
+2026-07-20 13:04:02,261   prompt 130/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.650  max_d_mean=1.36e-02
+2026-07-20 13:04:07,082   prompt 131/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.592  max_d_mean=1.32e-02
+2026-07-20 13:04:11,724   prompt 132/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.709  max_d_mean=1.35e-02
+2026-07-20 13:04:16,335   prompt 133/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.233  max_d_mean=1.77e-02
+2026-07-20 13:04:20,941   prompt 134/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.693  max_d_mean=1.26e-02
+2026-07-20 13:04:25,585   prompt 135/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.800  max_d_mean=1.41e-02
+2026-07-20 13:04:30,090   prompt 136/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.791  max_d_mean=1.42e-02
+2026-07-20 13:04:34,741   prompt 137/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.744  max_d_mean=1.40e-02
+2026-07-20 13:04:39,333   prompt 138/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.473  max_d_mean=1.96e-02
+2026-07-20 13:04:44,083   prompt 139/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.155  max_d_mean=1.65e-02
+2026-07-20 13:04:48,702   prompt 140/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.867  max_d_mean=1.44e-02
+2026-07-20 13:04:53,300   prompt 141/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.980  max_d_mean=1.41e-02
+2026-07-20 13:04:58,079   prompt 142/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.861  max_d_mean=1.37e-02
+2026-07-20 13:05:02,729   prompt 143/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.173  max_d_mean=1.58e-02
+2026-07-20 13:05:07,440   prompt 144/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.829  max_d_mean=1.31e-02
+2026-07-20 13:05:12,058   prompt 145/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.451  max_d_mean=1.78e-02
+2026-07-20 13:05:16,728   prompt 146/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.675  max_d_mean=1.21e-02
+2026-07-20 13:05:21,258   prompt 147/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.093  max_d_mean=1.47e-02
+2026-07-20 13:05:25,879   prompt 148/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.730  max_d_mean=1.15e-02
+2026-07-20 13:05:30,678   prompt 149/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.693  max_d_mean=1.15e-02
+2026-07-20 13:05:35,321   prompt 150/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.805  max_d_mean=2.00e-02
+2026-07-20 13:05:40,279   prompt 151/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.198  max_d_mean=1.52e-02
+2026-07-20 13:05:44,798   prompt 152/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.620  max_d_mean=1.97e-02
+2026-07-20 13:05:49,498   prompt 153/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.838  max_d_mean=1.22e-02
+2026-07-20 13:05:54,109   prompt 154/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.676  max_d_mean=1.89e-02
+2026-07-20 13:05:58,807   prompt 155/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.830  max_d_mean=1.17e-02
+2026-07-20 13:06:03,416   prompt 156/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.821  max_d_mean=1.26e-02
+2026-07-20 13:06:08,666   prompt 157/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.772  max_d_mean=1.24e-02
+2026-07-20 13:06:13,471   prompt 158/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.068  max_d_mean=1.41e-02
+2026-07-20 13:06:18,036   prompt 159/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.741  max_d_mean=1.15e-02
+2026-07-20 13:06:22,791   prompt 160/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.843  max_d_mean=1.23e-02
+2026-07-20 13:06:27,297   prompt 161/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.796  max_d_mean=1.21e-02
+2026-07-20 13:06:31,802   prompt 162/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.781  max_d_mean=1.16e-02
+2026-07-20 13:06:36,369   prompt 163/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.821  max_d_mean=1.16e-02
+2026-07-20 13:06:40,995   prompt 164/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.801  max_d_mean=1.11e-02
+2026-07-20 13:06:45,516   prompt 165/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.023  max_d_mean=1.23e-02
+2026-07-20 13:06:50,114   prompt 166/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.998  max_d_mean=1.31e-02
+2026-07-20 13:06:54,674   prompt 167/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.623  max_d_mean=1.68e-02
+2026-07-20 13:06:59,204   prompt 168/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.919  max_d_mean=1.19e-02
+2026-07-20 13:07:03,807   prompt 169/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.076  max_d_mean=1.31e-02
+2026-07-20 13:07:08,489   prompt 170/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.268  max_d_mean=1.42e-02
+2026-07-20 13:07:13,046   prompt 171/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.560  max_d_mean=9.07e-03
+2026-07-20 13:07:17,827   prompt 172/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.640  max_d_mean=9.13e-03
+2026-07-20 13:07:22,498   prompt 173/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.852  max_d_mean=1.13e-02
+2026-07-20 13:07:26,999   prompt 174/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.619  max_d_mean=1.02e-02
+2026-07-20 13:07:31,789   prompt 175/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.643  max_d_mean=9.73e-03
+2026-07-20 13:07:36,711   prompt 176/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.571  max_d_mean=9.13e-03
+2026-07-20 13:07:41,405   prompt 177/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.285  max_d_mean=1.34e-02
+2026-07-20 13:07:46,089   prompt 178/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.648  max_d_mean=9.53e-03
+2026-07-20 13:07:50,645   prompt 179/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.818  max_d_mean=1.03e-02
+2026-07-20 13:07:55,221   prompt 180/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.008  max_d_mean=1.15e-02
+2026-07-20 13:07:59,806   prompt 181/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.034  max_d_mean=1.18e-02
+2026-07-20 13:08:04,482   prompt 182/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=3.061  max_d_mean=1.78e-02
+2026-07-20 13:08:09,077   prompt 183/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.563  max_d_mean=1.49e-02
+2026-07-20 13:08:13,702   prompt 184/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.472  max_d_mean=1.46e-02
+2026-07-20 13:08:18,400   prompt 185/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.983  max_d_mean=1.09e-02
+2026-07-20 13:08:22,974   prompt 186/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.473  max_d_mean=1.46e-02
+2026-07-20 13:08:27,519   prompt 187/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.562  max_d_mean=8.20e-03
+2026-07-20 13:08:32,187   prompt 188/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.119  max_d_mean=1.19e-02
+2026-07-20 13:08:36,674   prompt 189/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=2.862  max_d_mean=1.63e-02
+2026-07-20 13:08:41,377   prompt 190/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.835  max_d_mean=1.04e-02
+2026-07-20 13:08:45,915   prompt 191/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.800  max_d_mean=1.04e-02
+2026-07-20 13:08:50,446   prompt 192/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.771  max_d_mean=9.78e-03
+2026-07-20 13:08:55,009   prompt 193/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.620  max_d_mean=8.17e-03
+2026-07-20 13:08:59,705   prompt 194/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.085  max_d_mean=1.16e-02
+2026-07-20 13:09:04,290   prompt 195/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.740  max_d_mean=1.50e-02
+2026-07-20 13:09:08,797   prompt 196/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.077  max_d_mean=1.12e-02
+2026-07-20 13:09:13,378   prompt 197/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.174  max_d_mean=1.23e-02
+2026-07-20 13:09:18,075   prompt 198/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=3.039  max_d_mean=1.67e-02
+2026-07-20 13:09:22,792   prompt 199/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.202  max_d_mean=1.16e-02
+2026-07-20 13:09:27,390   prompt 200/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.607  max_d_mean=7.41e-03
+2026-07-20 13:09:32,267   prompt 201/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.973  max_d_mean=1.03e-02
+2026-07-20 13:09:36,883   prompt 202/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.790  max_d_mean=9.37e-03
+2026-07-20 13:09:41,572   prompt 203/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.021  max_d_mean=1.05e-02
+2026-07-20 13:09:46,196   prompt 204/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.053  max_d_mean=1.07e-02
+2026-07-20 13:09:50,725   prompt 205/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.299  max_d_mean=1.22e-02
+2026-07-20 13:09:55,383   prompt 206/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.338  max_d_mean=1.24e-02
+2026-07-20 13:09:59,868   prompt 207/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=2.632  max_d_mean=1.35e-02
+2026-07-20 13:10:04,696   prompt 208/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.084  max_d_mean=1.06e-02
+2026-07-20 13:10:09,389   prompt 209/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.213  max_d_mean=1.10e-02
+2026-07-20 13:10:13,935   prompt 210/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=3.126  max_d_mean=1.60e-02
+2026-07-20 13:10:18,475   prompt 211/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.080  max_d_mean=1.04e-02
+2026-07-20 13:10:23,145   prompt 212/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.982  max_d_mean=9.31e-03
+2026-07-20 13:10:27,871   prompt 213/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.836  max_d_mean=8.78e-03
+2026-07-20 13:10:32,585   prompt 214/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.510  max_d_mean=6.32e-03
+2026-07-20 13:10:37,298   prompt 215/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.569  max_d_mean=7.78e-03
+2026-07-20 13:10:42,098   prompt 216/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.663  max_d_mean=7.60e-03
+2026-07-20 13:10:46,642   prompt 217/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.622  max_d_mean=5.70e-03
+2026-07-20 13:10:51,204   prompt 218/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.544  max_d_mean=7.30e-03
+2026-07-20 13:10:56,074   prompt 219/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.623  max_d_mean=7.90e-03
+2026-07-20 13:11:00,604   prompt 220/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.605  max_d_mean=7.56e-03
+2026-07-20 13:11:05,178   prompt 221/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.699  max_d_mean=8.35e-03
+2026-07-20 13:11:09,814   prompt 222/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.090  max_d_mean=1.01e-02
+2026-07-20 13:11:14,318   prompt 223/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.865  max_d_mean=8.71e-03
+2026-07-20 13:11:18,883   prompt 224/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.512  max_d_mean=7.04e-03
+2026-07-20 13:11:23,709   prompt 225/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.692  max_d_mean=8.00e-03
+2026-07-20 13:11:28,380   prompt 226/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.860  max_d_mean=8.71e-03
+2026-07-20 13:11:32,915   prompt 227/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.852  max_d_mean=8.52e-03
+2026-07-20 13:11:37,432   prompt 228/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.603  max_d_mean=7.62e-03
+2026-07-20 13:11:42,087   prompt 229/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.601  max_d_mean=6.70e-03
+2026-07-20 13:11:46,600   prompt 230/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.959  max_d_mean=8.94e-03
+2026-07-20 13:11:51,117   prompt 231/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.704  max_d_mean=8.13e-03
+2026-07-20 13:11:55,673   prompt 232/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.608  max_d_mean=7.04e-03
+2026-07-20 13:12:00,324   prompt 233/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.625  max_d_mean=7.65e-03
+2026-07-20 13:12:04,883   prompt 234/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.636  max_d_mean=7.07e-03
+2026-07-20 13:12:09,588   prompt 235/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.897  max_d_mean=8.73e-03
+2026-07-20 13:12:14,172   prompt 236/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.692  max_d_mean=7.33e-03
+2026-07-20 13:12:18,768   prompt 237/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.794  max_d_mean=7.86e-03
+2026-07-20 13:12:23,306   prompt 238/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.798  max_d_mean=7.65e-03
+2026-07-20 13:12:27,873   prompt 239/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.631  max_d_mean=7.05e-03
+2026-07-20 13:12:32,503   prompt 240/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.846  max_d_mean=8.18e-03
+2026-07-20 13:12:37,157   prompt 241/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.169  max_d_mean=9.62e-03
+2026-07-20 13:12:41,895   prompt 242/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.872  max_d_mean=8.12e-03
+2026-07-20 13:12:46,568   prompt 243/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.712  max_d_mean=6.32e-03
+2026-07-20 13:12:51,345   prompt 244/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.795  max_d_mean=7.76e-03
+2026-07-20 13:12:55,979   prompt 245/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.480  max_d_mean=6.38e-03
+2026-07-20 13:13:00,573   prompt 246/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.628  max_d_mean=5.71e-03
+2026-07-20 13:13:05,252   prompt 247/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.669  max_d_mean=7.14e-03
+2026-07-20 13:13:09,787   prompt 248/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.079  max_d_mean=8.80e-03
+2026-07-20 13:13:14,300   prompt 249/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.625  max_d_mean=6.34e-03
+2026-07-20 13:13:18,829   prompt 250/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.013  max_d_mean=8.69e-03
+2026-07-20 13:13:23,545   prompt 251/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.552  max_d_mean=6.30e-03
+2026-07-20 13:13:28,071   prompt 252/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.924  max_d_mean=8.05e-03
+2026-07-20 13:13:32,642   prompt 253/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.451  max_d_mean=1.04e-02
+2026-07-20 13:13:37,156   prompt 254/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.613  max_d_mean=6.83e-03
+2026-07-20 13:13:41,745   prompt 255/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.779  max_d_mean=7.41e-03
+2026-07-20 13:13:46,312   prompt 256/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.623  max_d_mean=6.88e-03
+2026-07-20 13:13:50,912   prompt 257/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.573  max_d_mean=5.74e-03
+2026-07-20 13:13:55,493   prompt 258/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.754  max_d_mean=7.06e-03
+2026-07-20 13:14:00,035   prompt 259/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.569  max_d_mean=6.27e-03
+2026-07-20 13:14:04,573   prompt 260/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.530  max_d_mean=1.07e-02
+2026-07-20 13:14:09,291   prompt 261/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.915  max_d_mean=7.99e-03
+2026-07-20 13:14:13,987   prompt 262/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.765  max_d_mean=7.39e-03
+2026-07-20 13:14:18,678   prompt 263/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.235  max_d_mean=9.40e-03
+2026-07-20 13:14:23,277   prompt 264/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.614  max_d_mean=1.08e-02
+2026-07-20 13:14:28,020   prompt 265/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.545  max_d_mean=1.08e-02
+2026-07-20 13:14:32,789   prompt 266/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.783  max_d_mean=1.15e-02
+2026-07-20 13:14:37,468   prompt 267/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=3.011  max_d_mean=1.22e-02
+2026-07-20 13:14:41,970   prompt 268/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.210  max_d_mean=8.99e-03
+2026-07-20 13:14:46,673   prompt 269/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.796  max_d_mean=7.04e-03
+2026-07-20 13:14:51,173   prompt 270/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.763  max_d_mean=7.31e-03
+2026-07-20 13:14:55,680   prompt 271/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.929  max_d_mean=7.70e-03
+2026-07-20 13:15:00,424   prompt 272/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.993  max_d_mean=8.09e-03
+2026-07-20 13:15:05,037   prompt 273/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.603  max_d_mean=5.44e-03
+2026-07-20 13:15:09,679   prompt 274/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.733  max_d_mean=6.17e-03
+2026-07-20 13:15:14,209   prompt 275/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.696  max_d_mean=5.67e-03
+2026-07-20 13:15:19,000   prompt 276/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.653  max_d_mean=5.69e-03
+2026-07-20 13:15:23,693   prompt 277/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.782  max_d_mean=7.06e-03
+2026-07-20 13:15:28,406   prompt 278/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.674  max_d_mean=6.40e-03
+2026-07-20 13:15:33,223   prompt 279/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.736  max_d_mean=5.61e-03
+2026-07-20 13:15:37,875   prompt 280/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.990  max_d_mean=7.63e-03
+2026-07-20 13:15:42,369   prompt 281/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=1.600  max_d_mean=5.38e-03
+2026-07-20 13:15:46,988   prompt 282/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.599  max_d_mean=5.03e-03
+2026-07-20 13:15:51,545   prompt 283/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.666  max_d_mean=5.70e-03
+2026-07-20 13:15:56,094   prompt 284/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.729  max_d_mean=6.39e-03
+2026-07-20 13:16:00,756   prompt 285/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.989  max_d_mean=7.17e-03
+2026-07-20 13:16:05,290   prompt 286/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.641  max_d_mean=5.76e-03
+2026-07-20 13:16:09,887   prompt 287/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.774  max_d_mean=5.30e-03
+2026-07-20 13:16:14,379   prompt 288/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=1.701  max_d_mean=6.42e-03
+2026-07-20 13:16:18,942   prompt 289/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.009  max_d_mean=7.49e-03
+2026-07-20 13:16:23,508   prompt 290/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.674  max_d_mean=6.07e-03
+2026-07-20 13:16:28,048   prompt 291/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.768  max_d_mean=5.53e-03
+2026-07-20 13:16:33,019   prompt 292/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.648  max_d_mean=5.35e-03
+2026-07-20 13:16:37,611   prompt 293/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.920  max_d_mean=6.96e-03
+2026-07-20 13:16:42,160   prompt 294/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.110  max_d_mean=7.82e-03
+2026-07-20 13:16:46,684   prompt 295/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.693  max_d_mean=4.77e-03
+2026-07-20 13:16:51,184   prompt 296/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=1.887  max_d_mean=6.76e-03
+2026-07-20 13:16:55,699   prompt 297/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.715  max_d_mean=6.31e-03
+2026-07-20 13:17:00,272   prompt 298/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.765  max_d_mean=5.41e-03
+2026-07-20 13:17:04,892   prompt 299/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.816  max_d_mean=6.82e-03
+2026-07-20 13:17:09,478   prompt 300/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.735  max_d_mean=5.70e-03
+2026-07-20 13:17:14,188   prompt 301/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.910  max_d_mean=6.55e-03
+2026-07-20 13:17:18,801   prompt 302/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.813  max_d_mean=5.81e-03
+2026-07-20 13:17:23,284   prompt 303/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=1.835  max_d_mean=6.11e-03
+2026-07-20 13:17:27,801   prompt 304/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.778  max_d_mean=6.23e-03
+2026-07-20 13:17:32,298   prompt 305/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=1.634  max_d_mean=5.33e-03
+2026-07-20 13:17:36,795   prompt 306/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=1.987  max_d_mean=7.01e-03
+2026-07-20 13:17:41,294   prompt 307/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=1.843  max_d_mean=6.27e-03
+2026-07-20 13:17:46,079   prompt 308/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.878  max_d_mean=6.55e-03
+2026-07-20 13:17:50,808   prompt 309/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.707  max_d_mean=4.98e-03
+2026-07-20 13:17:55,351   prompt 310/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.913  max_d_mean=6.54e-03
+2026-07-20 13:18:00,006   prompt 311/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.767  max_d_mean=5.91e-03
+2026-07-20 13:18:04,559   prompt 312/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.993  max_d_mean=6.83e-03
+2026-07-20 13:18:09,103   prompt 313/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.712  max_d_mean=5.29e-03
+2026-07-20 13:18:13,706   prompt 314/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.696  max_d_mean=5.48e-03
+2026-07-20 13:18:18,411   prompt 315/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.878  max_d_mean=6.53e-03
+2026-07-20 13:18:23,017   prompt 316/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.167  max_d_mean=7.29e-03
+2026-07-20 13:18:27,563   prompt 317/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.639  max_d_mean=5.39e-03
+2026-07-20 13:18:32,183   prompt 318/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.043  max_d_mean=6.97e-03
+2026-07-20 13:18:36,718   prompt 319/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.968  max_d_mean=6.46e-03
+2026-07-20 13:18:41,316   prompt 320/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.991  max_d_mean=6.67e-03
+2026-07-20 13:18:45,985   prompt 321/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.936  max_d_mean=6.20e-03
+2026-07-20 13:18:50,810   prompt 322/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.949  max_d_mean=6.59e-03
+2026-07-20 13:18:55,282   prompt 323/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=1.773  max_d_mean=5.50e-03
+2026-07-20 13:18:59,966   prompt 324/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.583  max_d_mean=5.18e-03
+2026-07-20 13:19:04,631   prompt 325/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.748  max_d_mean=5.36e-03
+2026-07-20 13:19:09,385   prompt 326/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.886  max_d_mean=6.24e-03
+2026-07-20 13:19:13,905   prompt 327/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.727  max_d_mean=4.97e-03
+2026-07-20 13:19:18,476   prompt 328/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.698  max_d_mean=5.08e-03
+2026-07-20 13:19:22,965   prompt 329/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=1.725  max_d_mean=4.69e-03
+2026-07-20 13:19:27,585   prompt 330/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.664  max_d_mean=5.13e-03
+2026-07-20 13:19:32,185   prompt 331/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.674  max_d_mean=5.08e-03
+2026-07-20 13:19:36,875   prompt 332/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.811  max_d_mean=5.84e-03
+2026-07-20 13:19:41,378   prompt 333/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.729  max_d_mean=5.23e-03
+2026-07-20 13:19:45,896   prompt 334/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.607  max_d_mean=4.65e-03
+2026-07-20 13:19:50,390   prompt 335/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=1.765  max_d_mean=5.80e-03
+2026-07-20 13:19:55,168   prompt 336/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.002  max_d_mean=6.30e-03
+2026-07-20 13:19:59,662   prompt 337/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=1.828  max_d_mean=5.52e-03
+2026-07-20 13:20:04,284   prompt 338/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.759  max_d_mean=5.12e-03
+2026-07-20 13:20:08,774   prompt 339/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=1.955  max_d_mean=6.07e-03
+2026-07-20 13:20:13,509   prompt 340/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.916  max_d_mean=5.95e-03
+2026-07-20 13:20:18,111   prompt 341/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.744  max_d_mean=5.20e-03
+2026-07-20 13:20:22,687   prompt 342/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.789  max_d_mean=4.92e-03
+2026-07-20 13:20:27,196   prompt 343/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.817  max_d_mean=5.49e-03
+2026-07-20 13:20:31,867   prompt 344/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.673  max_d_mean=4.48e-03
+2026-07-20 13:20:36,476   prompt 345/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.758  max_d_mean=4.47e-03
+2026-07-20 13:20:41,050   prompt 346/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.928  max_d_mean=9.33e-03
+2026-07-20 13:20:45,687   prompt 347/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.941  max_d_mean=6.26e-03
+2026-07-20 13:20:50,254   prompt 348/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.464  max_d_mean=8.06e-03
+2026-07-20 13:20:55,075   prompt 349/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.756  max_d_mean=5.13e-03
+2026-07-20 13:20:59,702   prompt 350/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.793  max_d_mean=5.26e-03
+2026-07-20 13:21:04,402   prompt 351/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.050  max_d_mean=6.15e-03
+2026-07-20 13:21:09,110   prompt 352/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.818  max_d_mean=5.52e-03
+2026-07-20 13:21:13,628   prompt 353/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.886  max_d_mean=5.96e-03
+2026-07-20 13:21:18,290   prompt 354/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.705  max_d_mean=4.53e-03
+2026-07-20 13:21:23,000   prompt 355/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.669  max_d_mean=4.17e-03
+2026-07-20 13:21:27,512   prompt 356/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.734  max_d_mean=4.34e-03
+2026-07-20 13:21:32,089   prompt 357/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.667  max_d_mean=4.21e-03
+2026-07-20 13:21:36,614   prompt 358/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.675  max_d_mean=4.73e-03
+2026-07-20 13:21:41,178   prompt 359/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.748  max_d_mean=4.91e-03
+2026-07-20 13:21:45,794   prompt 360/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.688  max_d_mean=4.95e-03
+2026-07-20 13:21:50,392   prompt 361/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.682  max_d_mean=3.80e-03
+2026-07-20 13:21:54,908   prompt 362/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.573  max_d_mean=4.22e-03
+2026-07-20 13:21:59,591   prompt 363/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.617  max_d_mean=4.61e-03
+2026-07-20 13:22:04,104   prompt 364/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.717  max_d_mean=4.90e-03
+2026-07-20 13:22:08,703   prompt 365/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.918  max_d_mean=5.67e-03
+2026-07-20 13:22:13,509   prompt 366/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.793  max_d_mean=5.24e-03
+2026-07-20 13:22:18,010   prompt 367/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.744  max_d_mean=4.96e-03
+2026-07-20 13:22:22,686   prompt 368/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.058  max_d_mean=6.08e-03
+2026-07-20 13:22:27,284   prompt 369/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.211  max_d_mean=6.56e-03
+2026-07-20 13:22:31,774   prompt 370/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=1.654  max_d_mean=4.81e-03
+2026-07-20 13:22:36,489   prompt 371/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.648  max_d_mean=4.53e-03
+2026-07-20 13:22:41,104   prompt 372/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.626  max_d_mean=4.30e-03
+2026-07-20 13:22:46,279   prompt 373/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.924  max_d_mean=5.63e-03
+2026-07-20 13:22:50,972   prompt 374/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.128  max_d_mean=6.37e-03
+2026-07-20 13:22:55,812   prompt 375/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.124  max_d_mean=6.13e-03
+2026-07-20 13:23:00,489   prompt 376/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=1.575  max_d_mean=4.39e-03
+2026-07-20 13:23:05,088   prompt 377/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.840  max_d_mean=5.49e-03
+2026-07-20 13:23:09,846   prompt 378/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.664  max_d_mean=4.32e-03
+2026-07-20 13:23:14,589   prompt 379/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.741  max_d_mean=5.04e-03
+2026-07-20 13:23:19,090   prompt 380/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.804  max_d_mean=5.37e-03
+2026-07-20 13:23:23,726   prompt 381/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.550  max_d_mean=7.58e-03
+2026-07-20 13:23:28,279   prompt 382/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.981  max_d_mean=5.58e-03
+2026-07-20 13:23:32,894   prompt 383/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.112  max_d_mean=6.04e-03
+2026-07-20 13:23:37,573   prompt 384/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.372  max_d_mean=6.86e-03
+2026-07-20 13:23:42,111   prompt 385/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.062  max_d_mean=6.01e-03
+2026-07-20 13:23:46,750   prompt 386/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.762  max_d_mean=4.99e-03
+2026-07-20 13:23:51,703   prompt 387/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.208  max_d_mean=6.13e-03
+2026-07-20 13:23:56,387   prompt 388/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.378  max_d_mean=6.67e-03
+2026-07-20 13:24:01,094   prompt 389/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.704  max_d_mean=4.59e-03
+2026-07-20 13:24:05,717   prompt 390/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.004  max_d_mean=5.47e-03
+2026-07-20 13:24:10,507   prompt 391/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.714  max_d_mean=4.93e-03
+2026-07-20 13:24:15,063   prompt 392/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.953  max_d_mean=5.44e-03
+2026-07-20 13:24:19,710   prompt 393/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.824  max_d_mean=4.90e-03
+2026-07-20 13:24:24,264   prompt 394/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.751  max_d_mean=4.45e-03
+2026-07-20 13:24:28,900   prompt 395/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.614  max_d_mean=4.39e-03
+2026-07-20 13:24:33,378   prompt 396/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=2.269  max_d_mean=6.04e-03
+2026-07-20 13:24:38,094   prompt 397/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.665  max_d_mean=4.51e-03
+2026-07-20 13:24:42,788   prompt 398/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.307  max_d_mean=6.08e-03
+2026-07-20 13:24:47,414   prompt 399/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.959  max_d_mean=5.25e-03
+2026-07-20 13:24:52,095   prompt 400/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.240  max_d_mean=6.01e-03
+2026-07-20 13:24:56,812   prompt 401/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.101  max_d_mean=5.62e-03
+2026-07-20 13:25:01,447   prompt 402/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.698  max_d_mean=4.21e-03
+2026-07-20 13:25:05,999   prompt 403/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.124  max_d_mean=5.76e-03
+2026-07-20 13:25:10,599   prompt 404/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.677  max_d_mean=4.42e-03
+2026-07-20 13:25:15,599   prompt 405/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.206  max_d_mean=6.04e-03
+2026-07-20 13:25:20,265   prompt 406/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.189  max_d_mean=5.75e-03
+2026-07-20 13:25:24,756   prompt 407/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=2.212  max_d_mean=5.86e-03
+2026-07-20 13:25:29,511   prompt 408/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.735  max_d_mean=4.46e-03
+2026-07-20 13:25:34,180   prompt 409/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.708  max_d_mean=4.24e-03
+2026-07-20 13:25:38,686   prompt 410/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.810  max_d_mean=4.73e-03
+2026-07-20 13:25:43,292   prompt 411/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.830  max_d_mean=4.75e-03
+2026-07-20 13:25:47,777   prompt 412/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=2.035  max_d_mean=5.51e-03
+2026-07-20 13:25:52,494   prompt 413/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.862  max_d_mean=4.83e-03
+2026-07-20 13:25:57,780   prompt 414/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.695  max_d_mean=4.02e-03
+2026-07-20 13:26:02,404   prompt 415/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.114  max_d_mean=5.77e-03
+2026-07-20 13:26:07,110   prompt 416/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.787  max_d_mean=4.55e-03
+2026-07-20 13:26:11,597   prompt 417/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=2.589  max_d_mean=6.79e-03
+2026-07-20 13:26:16,201   prompt 418/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.001  max_d_mean=5.04e-03
+2026-07-20 13:26:20,844   prompt 419/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.275  max_d_mean=5.78e-03
+2026-07-20 13:26:25,720   prompt 420/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.213  max_d_mean=5.51e-03
+2026-07-20 13:26:30,212   prompt 421/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=2.050  max_d_mean=5.12e-03
+2026-07-20 13:26:34,799   prompt 422/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.010  max_d_mean=4.77e-03
+2026-07-20 13:26:39,513   prompt 423/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.538  max_d_mean=6.60e-03
+2026-07-20 13:26:44,240   prompt 424/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.920  max_d_mean=4.97e-03
+2026-07-20 13:26:49,067   prompt 425/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.731  max_d_mean=4.29e-03
+2026-07-20 13:26:53,845   prompt 426/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.179  max_d_mean=5.55e-03
+2026-07-20 13:26:58,493   prompt 427/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.749  max_d_mean=4.32e-03
+2026-07-20 13:27:02,998   prompt 428/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.672  max_d_mean=4.25e-03
+2026-07-20 13:27:07,787   prompt 429/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.696  max_d_mean=4.14e-03
+2026-07-20 13:27:12,420   prompt 430/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.941  max_d_mean=4.74e-03
+2026-07-20 13:27:16,989   prompt 431/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.742  max_d_mean=4.16e-03
+2026-07-20 13:27:21,510   prompt 432/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.761  max_d_mean=4.19e-03
+2026-07-20 13:27:26,571   prompt 433/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.738  max_d_mean=4.14e-03
+2026-07-20 13:27:31,218   prompt 434/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.764  max_d_mean=4.28e-03
+2026-07-20 13:27:35,932   prompt 435/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.010  max_d_mean=4.85e-03
+2026-07-20 13:27:40,581   prompt 436/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.901  max_d_mean=4.51e-03
+2026-07-20 13:27:45,081   prompt 437/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=2.055  max_d_mean=4.91e-03
+2026-07-20 13:27:49,612   prompt 438/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.753  max_d_mean=4.48e-03
+2026-07-20 13:27:54,195   prompt 439/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.921  max_d_mean=4.69e-03
+2026-07-20 13:27:58,838   prompt 440/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.800  max_d_mean=4.38e-03
+2026-07-20 13:28:03,600   prompt 441/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.732  max_d_mean=4.29e-03
+2026-07-20 13:28:08,110   prompt 442/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.123  max_d_mean=5.18e-03
+2026-07-20 13:28:12,691   prompt 443/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.109  max_d_mean=5.13e-03
+2026-07-20 13:28:17,628   prompt 444/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=3.469  max_d_mean=8.96e-03
+2026-07-20 13:28:22,367   prompt 445/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.378  max_d_mean=5.78e-03
+2026-07-20 13:28:27,196   prompt 446/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.175  max_d_mean=5.55e-03
+2026-07-20 13:28:31,838   prompt 447/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.085  max_d_mean=4.96e-03
+2026-07-20 13:28:36,352   prompt 448/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.064  max_d_mean=4.93e-03
+2026-07-20 13:28:40,851   prompt 449/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=2.527  max_d_mean=6.27e-03
+2026-07-20 13:28:45,392   prompt 450/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.453  max_d_mean=6.01e-03
+2026-07-20 13:28:50,026   prompt 451/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=2.360  max_d_mean=5.86e-03
+2026-07-20 13:28:54,693   prompt 452/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.099  max_d_mean=5.03e-03
+2026-07-20 13:28:59,730   prompt 453/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.539  max_d_mean=3.68e-03
+2026-07-20 13:29:04,421   prompt 454/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.766  max_d_mean=4.13e-03
+2026-07-20 13:29:09,389   prompt 455/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.960  max_d_mean=4.59e-03
+2026-07-20 13:29:14,716   prompt 456/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.723  max_d_mean=4.13e-03
+2026-07-20 13:29:19,444   prompt 457/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.659  max_d_mean=3.76e-03
+2026-07-20 13:29:24,249   prompt 458/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.784  max_d_mean=4.14e-03
+2026-07-20 13:29:28,982   prompt 459/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.613  max_d_mean=3.68e-03
+2026-07-20 13:29:33,567   prompt 460/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.130  max_d_mean=4.87e-03
+2026-07-20 13:29:38,246   prompt 461/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.735  max_d_mean=3.97e-03
+2026-07-20 13:29:43,139   prompt 462/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.694  max_d_mean=3.83e-03
+2026-07-20 13:29:48,080   prompt 463/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.768  max_d_mean=4.25e-03
+2026-07-20 13:29:52,681   prompt 464/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.691  max_d_mean=3.90e-03
+2026-07-20 13:29:57,211   prompt 465/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.372  max_d_mean=5.62e-03
+2026-07-20 13:30:01,781   prompt 466/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.157  max_d_mean=5.07e-03
+2026-07-20 13:30:06,418   prompt 467/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.536  max_d_mean=3.68e-03
+2026-07-20 13:30:10,961   prompt 468/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.333  max_d_mean=5.55e-03
+2026-07-20 13:30:15,511   prompt 469/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.062  max_d_mean=4.74e-03
+2026-07-20 13:30:20,080   prompt 470/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.356  max_d_mean=5.48e-03
+2026-07-20 13:30:24,884   prompt 471/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.026  max_d_mean=4.57e-03
+2026-07-20 13:30:30,012   prompt 472/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.857  max_d_mean=6.79e-03
+2026-07-20 13:30:35,991   prompt 473/1000  seq_len=128 n_valid=111  6s  max||J||/sqrt(d)=1.989  max_d_mean=4.78e-03
+2026-07-20 13:30:40,802   prompt 474/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.409  max_d_mean=5.71e-03
+2026-07-20 13:30:45,296   prompt 475/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=2.139  max_d_mean=4.90e-03
+2026-07-20 13:30:50,027   prompt 476/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.753  max_d_mean=3.91e-03
+2026-07-20 13:30:54,624   prompt 477/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.618  max_d_mean=3.75e-03
+2026-07-20 13:30:59,207   prompt 478/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.800  max_d_mean=4.10e-03
+2026-07-20 13:31:03,874   prompt 479/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.315  max_d_mean=5.26e-03
+2026-07-20 13:31:08,490   prompt 480/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.629  max_d_mean=3.39e-03
+2026-07-20 13:31:13,067   prompt 481/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.733  max_d_mean=3.95e-03
+2026-07-20 13:31:17,733   prompt 482/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.781  max_d_mean=4.08e-03
+2026-07-20 13:31:22,312   prompt 483/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.942  max_d_mean=4.37e-03
+2026-07-20 13:31:27,267   prompt 484/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.657  max_d_mean=3.77e-03
+2026-07-20 13:31:31,890   prompt 485/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.933  max_d_mean=4.37e-03
+2026-07-20 13:31:36,468   prompt 486/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.826  max_d_mean=3.98e-03
+2026-07-20 13:31:41,107   prompt 487/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.916  max_d_mean=4.35e-03
+2026-07-20 13:31:45,973   prompt 488/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.751  max_d_mean=3.87e-03
+2026-07-20 13:31:50,666   prompt 489/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.710  max_d_mean=3.88e-03
+2026-07-20 13:31:55,279   prompt 490/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.665  max_d_mean=3.76e-03
+2026-07-20 13:32:00,022   prompt 491/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.714  max_d_mean=3.69e-03
+2026-07-20 13:32:04,716   prompt 492/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.374  max_d_mean=5.20e-03
+2026-07-20 13:32:09,400   prompt 493/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.947  max_d_mean=4.42e-03
+2026-07-20 13:32:14,059   prompt 494/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.837  max_d_mean=4.19e-03
+2026-07-20 13:32:19,284   prompt 495/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.084  max_d_mean=4.52e-03
+2026-07-20 13:32:24,863   prompt 496/1000  seq_len=128 n_valid=111  6s  max||J||/sqrt(d)=2.404  max_d_mean=5.41e-03
+2026-07-20 13:32:30,295   prompt 497/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.595  max_d_mean=5.68e-03
+2026-07-20 13:32:35,869   prompt 498/1000  seq_len=128 n_valid=111  6s  max||J||/sqrt(d)=1.804  max_d_mean=3.87e-03
+2026-07-20 13:32:40,682   prompt 499/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.126  max_d_mean=4.52e-03
+2026-07-20 13:32:45,417   prompt 500/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.848  max_d_mean=4.11e-03
+2026-07-20 13:32:50,205   prompt 501/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.883  max_d_mean=3.92e-03
+2026-07-20 13:32:54,945   prompt 502/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.110  max_d_mean=4.52e-03
+2026-07-20 13:32:59,470   prompt 503/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.998  max_d_mean=4.34e-03
+2026-07-20 13:33:04,105   prompt 504/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.738  max_d_mean=3.68e-03
+2026-07-20 13:33:08,757   prompt 505/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.087  max_d_mean=4.54e-03
+2026-07-20 13:33:13,386   prompt 506/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.584  max_d_mean=5.53e-03
+2026-07-20 13:33:18,003   prompt 507/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.030  max_d_mean=4.43e-03
+2026-07-20 13:33:22,880   prompt 508/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.970  max_d_mean=4.10e-03
+2026-07-20 13:33:27,514   prompt 509/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.986  max_d_mean=4.30e-03
+2026-07-20 13:33:32,185   prompt 510/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.992  max_d_mean=4.31e-03
+2026-07-20 13:33:36,886   prompt 511/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.010  max_d_mean=4.20e-03
+2026-07-20 13:33:41,395   prompt 512/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.685  max_d_mean=3.38e-03
+2026-07-20 13:33:45,905   prompt 513/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.012  max_d_mean=4.22e-03
+2026-07-20 13:33:50,423   prompt 514/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.228  max_d_mean=4.70e-03
+2026-07-20 13:33:54,938   prompt 515/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.052  max_d_mean=4.29e-03
+2026-07-20 13:33:59,569   prompt 516/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.002  max_d_mean=4.04e-03
+2026-07-20 13:34:04,088   prompt 517/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.936  max_d_mean=3.99e-03
+2026-07-20 13:34:08,596   prompt 518/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.784  max_d_mean=3.75e-03
+2026-07-20 13:34:13,420   prompt 519/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.648  max_d_mean=5.61e-03
+2026-07-20 13:34:18,018   prompt 520/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.197  max_d_mean=4.51e-03
+2026-07-20 13:34:22,543   prompt 521/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.964  max_d_mean=4.13e-03
+2026-07-20 13:34:27,105   prompt 522/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.101  max_d_mean=4.43e-03
+2026-07-20 13:34:31,686   prompt 523/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.555  max_d_mean=3.31e-03
+2026-07-20 13:34:36,284   prompt 524/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.790  max_d_mean=3.65e-03
+2026-07-20 13:34:40,937   prompt 525/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.291  max_d_mean=4.90e-03
+2026-07-20 13:34:45,707   prompt 526/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.821  max_d_mean=3.92e-03
+2026-07-20 13:34:50,272   prompt 527/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.168  max_d_mean=4.59e-03
+2026-07-20 13:34:54,867   prompt 528/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.741  max_d_mean=5.74e-03
+2026-07-20 13:34:59,582   prompt 529/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.894  max_d_mean=3.78e-03
+2026-07-20 13:35:04,292   prompt 530/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.339  max_d_mean=4.83e-03
+2026-07-20 13:35:09,015   prompt 531/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.177  max_d_mean=4.36e-03
+2026-07-20 13:35:13,646   prompt 532/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.266  max_d_mean=4.61e-03
+2026-07-20 13:35:18,166   prompt 533/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.207  max_d_mean=4.45e-03
+2026-07-20 13:35:22,743   prompt 534/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.921  max_d_mean=3.89e-03
+2026-07-20 13:35:27,394   prompt 535/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.932  max_d_mean=3.97e-03
+2026-07-20 13:35:31,937   prompt 536/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.233  max_d_mean=4.49e-03
+2026-07-20 13:35:36,690   prompt 537/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.804  max_d_mean=3.74e-03
+2026-07-20 13:35:41,293   prompt 538/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.654  max_d_mean=3.40e-03
+2026-07-20 13:35:45,913   prompt 539/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.032  max_d_mean=4.03e-03
+2026-07-20 13:35:50,617   prompt 540/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.939  max_d_mean=4.00e-03
+2026-07-20 13:35:55,219   prompt 541/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.958  max_d_mean=3.98e-03
+2026-07-20 13:35:59,778   prompt 542/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.881  max_d_mean=3.76e-03
+2026-07-20 13:36:04,265   prompt 543/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=1.964  max_d_mean=3.90e-03
+2026-07-20 13:36:08,910   prompt 544/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.542  max_d_mean=2.95e-03
+2026-07-20 13:36:13,501   prompt 545/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.580  max_d_mean=3.17e-03
+2026-07-20 13:36:18,110   prompt 546/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.631  max_d_mean=3.15e-03
+2026-07-20 13:36:22,732   prompt 547/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.766  max_d_mean=3.44e-03
+2026-07-20 13:36:27,401   prompt 548/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.634  max_d_mean=3.25e-03
+2026-07-20 13:36:32,168   prompt 549/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.824  max_d_mean=3.45e-03
+2026-07-20 13:36:37,093   prompt 550/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.829  max_d_mean=3.58e-03
+2026-07-20 13:36:42,035   prompt 551/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.922  max_d_mean=3.85e-03
+2026-07-20 13:36:46,594   prompt 552/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.575  max_d_mean=3.16e-03
+2026-07-20 13:36:51,167   prompt 553/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.024  max_d_mean=3.83e-03
+2026-07-20 13:36:55,853   prompt 554/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.687  max_d_mean=3.24e-03
+2026-07-20 13:37:00,369   prompt 555/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.768  max_d_mean=3.46e-03
+2026-07-20 13:37:05,013   prompt 556/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.879  max_d_mean=3.72e-03
+2026-07-20 13:37:09,680   prompt 557/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.630  max_d_mean=2.76e-03
+2026-07-20 13:37:14,294   prompt 558/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.962  max_d_mean=3.73e-03
+2026-07-20 13:37:18,789   prompt 559/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=1.991  max_d_mean=3.82e-03
+2026-07-20 13:37:23,309   prompt 560/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.598  max_d_mean=2.93e-03
+2026-07-20 13:37:27,886   prompt 561/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.992  max_d_mean=3.74e-03
+2026-07-20 13:37:32,614   prompt 562/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.815  max_d_mean=3.46e-03
+2026-07-20 13:37:37,411   prompt 563/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.656  max_d_mean=3.00e-03
+2026-07-20 13:37:41,907   prompt 564/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=1.816  max_d_mean=3.36e-03
+2026-07-20 13:37:46,606   prompt 565/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.551  max_d_mean=2.77e-03
+2026-07-20 13:37:51,299   prompt 566/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.115  max_d_mean=4.17e-03
+2026-07-20 13:37:55,871   prompt 567/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.799  max_d_mean=3.43e-03
+2026-07-20 13:38:00,473   prompt 568/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.764  max_d_mean=3.48e-03
+2026-07-20 13:38:05,291   prompt 569/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.628  max_d_mean=2.93e-03
+2026-07-20 13:38:10,026   prompt 570/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.287  max_d_mean=4.40e-03
+2026-07-20 13:38:14,799   prompt 571/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.792  max_d_mean=3.28e-03
+2026-07-20 13:38:19,546   prompt 572/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.766  max_d_mean=3.47e-03
+2026-07-20 13:38:24,112   prompt 573/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.749  max_d_mean=3.41e-03
+2026-07-20 13:38:28,635   prompt 574/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.118  max_d_mean=4.22e-03
+2026-07-20 13:38:33,206   prompt 575/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.038  max_d_mean=3.83e-03
+2026-07-20 13:38:37,895   prompt 576/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.691  max_d_mean=3.09e-03
+2026-07-20 13:38:42,399   prompt 577/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.830  max_d_mean=3.44e-03
+2026-07-20 13:38:46,904   prompt 578/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.748  max_d_mean=3.32e-03
+2026-07-20 13:38:51,409   prompt 579/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.338  max_d_mean=4.43e-03
+2026-07-20 13:38:55,926   prompt 580/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.645  max_d_mean=2.93e-03
+2026-07-20 13:39:00,480   prompt 581/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.879  max_d_mean=3.69e-03
+2026-07-20 13:39:04,985   prompt 582/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.070  max_d_mean=3.95e-03
+2026-07-20 13:39:09,487   prompt 583/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.924  max_d_mean=3.60e-03
+2026-07-20 13:39:14,146   prompt 584/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.685  max_d_mean=5.31e-03
+2026-07-20 13:39:18,666   prompt 585/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.217  max_d_mean=4.13e-03
+2026-07-20 13:39:23,209   prompt 586/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.066  max_d_mean=3.82e-03
+2026-07-20 13:39:27,921   prompt 587/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.646  max_d_mean=3.14e-03
+2026-07-20 13:39:32,516   prompt 588/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.822  max_d_mean=3.47e-03
+2026-07-20 13:39:37,201   prompt 589/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.770  max_d_mean=3.21e-03
+2026-07-20 13:39:41,987   prompt 590/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.971  max_d_mean=3.53e-03
+2026-07-20 13:39:46,568   prompt 591/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.521  max_d_mean=2.79e-03
+2026-07-20 13:39:51,066   prompt 592/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=2.015  max_d_mean=3.84e-03
+2026-07-20 13:39:55,676   prompt 593/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.322  max_d_mean=4.38e-03
+2026-07-20 13:40:00,239   prompt 594/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.650  max_d_mean=3.04e-03
+2026-07-20 13:40:04,741   prompt 595/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.922  max_d_mean=3.69e-03
+2026-07-20 13:40:09,255   prompt 596/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.927  max_d_mean=3.53e-03
+2026-07-20 13:40:13,755   prompt 597/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.657  max_d_mean=2.85e-03
+2026-07-20 13:40:18,335   prompt 598/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.079  max_d_mean=3.73e-03
+2026-07-20 13:40:22,890   prompt 599/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.948  max_d_mean=3.46e-03
+2026-07-20 13:40:27,380   prompt 600/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=2.145  max_d_mean=3.87e-03
+2026-07-20 13:40:32,090   prompt 601/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.942  max_d_mean=3.58e-03
+2026-07-20 13:40:36,596   prompt 602/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.008  max_d_mean=3.59e-03
+2026-07-20 13:40:41,433   prompt 603/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.784  max_d_mean=2.81e-03
+2026-07-20 13:40:45,923   prompt 604/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=1.804  max_d_mean=3.32e-03
+2026-07-20 13:40:50,504   prompt 605/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.978  max_d_mean=3.50e-03
+2026-07-20 13:40:55,025   prompt 606/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.765  max_d_mean=3.15e-03
+2026-07-20 13:40:59,565   prompt 607/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.903  max_d_mean=3.28e-03
+2026-07-20 13:41:04,075   prompt 608/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.694  max_d_mean=3.03e-03
+2026-07-20 13:41:08,726   prompt 609/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.334  max_d_mean=4.32e-03
+2026-07-20 13:41:13,485   prompt 610/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.731  max_d_mean=3.05e-03
+2026-07-20 13:41:18,129   prompt 611/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.170  max_d_mean=3.81e-03
+2026-07-20 13:41:22,635   prompt 612/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.663  max_d_mean=2.61e-03
+2026-07-20 13:41:27,166   prompt 613/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.610  max_d_mean=2.65e-03
+2026-07-20 13:41:31,867   prompt 614/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.668  max_d_mean=2.93e-03
+2026-07-20 13:41:36,514   prompt 615/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.402  max_d_mean=4.46e-03
+2026-07-20 13:41:41,138   prompt 616/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.123  max_d_mean=3.72e-03
+2026-07-20 13:41:45,681   prompt 617/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.554  max_d_mean=2.54e-03
+2026-07-20 13:41:50,339   prompt 618/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.723  max_d_mean=3.05e-03
+2026-07-20 13:41:55,005   prompt 619/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.514  max_d_mean=2.25e-03
+2026-07-20 13:41:59,691   prompt 620/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.486  max_d_mean=2.41e-03
+2026-07-20 13:42:04,516   prompt 621/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.731  max_d_mean=2.93e-03
+2026-07-20 13:42:09,169   prompt 622/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.565  max_d_mean=2.81e-03
+2026-07-20 13:42:13,887   prompt 623/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.521  max_d_mean=2.49e-03
+2026-07-20 13:42:18,383   prompt 624/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=2.471  max_d_mean=4.31e-03
+2026-07-20 13:42:23,101   prompt 625/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.591  max_d_mean=2.79e-03
+2026-07-20 13:42:27,804   prompt 626/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.470  max_d_mean=2.31e-03
+2026-07-20 13:42:32,426   prompt 627/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.104  max_d_mean=3.60e-03
+2026-07-20 13:42:37,086   prompt 628/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.840  max_d_mean=3.25e-03
+2026-07-20 13:42:41,783   prompt 629/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.315  max_d_mean=4.00e-03
+2026-07-20 13:42:46,627   prompt 630/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.743  max_d_mean=3.03e-03
+2026-07-20 13:42:51,699   prompt 631/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.610  max_d_mean=2.69e-03
+2026-07-20 13:42:56,277   prompt 632/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.921  max_d_mean=3.25e-03
+2026-07-20 13:43:00,799   prompt 633/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.538  max_d_mean=2.36e-03
+2026-07-20 13:43:05,383   prompt 634/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.663  max_d_mean=2.92e-03
+2026-07-20 13:43:10,114   prompt 635/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.708  max_d_mean=2.90e-03
+2026-07-20 13:43:14,694   prompt 636/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.837  max_d_mean=3.08e-03
+2026-07-20 13:43:19,180   prompt 637/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=1.862  max_d_mean=3.19e-03
+2026-07-20 13:43:23,704   prompt 638/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.788  max_d_mean=2.99e-03
+2026-07-20 13:43:28,416   prompt 639/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.666  max_d_mean=2.74e-03
+2026-07-20 13:43:32,973   prompt 640/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.656  max_d_mean=2.79e-03
+2026-07-20 13:43:37,723   prompt 641/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.687  max_d_mean=2.88e-03
+2026-07-20 13:43:42,419   prompt 642/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.788  max_d_mean=3.03e-03
+2026-07-20 13:43:47,197   prompt 643/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.928  max_d_mean=3.33e-03
+2026-07-20 13:43:51,895   prompt 644/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.023  max_d_mean=3.36e-03
+2026-07-20 13:43:56,389   prompt 645/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=2.002  max_d_mean=3.46e-03
+2026-07-20 13:44:01,010   prompt 646/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.831  max_d_mean=3.14e-03
+2026-07-20 13:44:05,609   prompt 647/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.694  max_d_mean=2.79e-03
+2026-07-20 13:44:10,610   prompt 648/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.312  max_d_mean=3.83e-03
+2026-07-20 13:44:15,135   prompt 649/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.752  max_d_mean=2.97e-03
+2026-07-20 13:44:20,023   prompt 650/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.709  max_d_mean=2.81e-03
+2026-07-20 13:44:24,717   prompt 651/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.471  max_d_mean=2.27e-03
+2026-07-20 13:44:29,358   prompt 652/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.631  max_d_mean=2.76e-03
+2026-07-20 13:44:33,925   prompt 653/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.913  max_d_mean=3.23e-03
+2026-07-20 13:44:38,618   prompt 654/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.541  max_d_mean=2.66e-03
+2026-07-20 13:44:43,188   prompt 655/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.924  max_d_mean=3.23e-03
+2026-07-20 13:44:47,752   prompt 656/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.762  max_d_mean=3.02e-03
+2026-07-20 13:44:52,313   prompt 657/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.516  max_d_mean=2.30e-03
+2026-07-20 13:44:57,170   prompt 658/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.971  max_d_mean=3.15e-03
+2026-07-20 13:45:01,892   prompt 659/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.711  max_d_mean=2.80e-03
+2026-07-20 13:45:06,490   prompt 660/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.619  max_d_mean=2.67e-03
+2026-07-20 13:45:10,999   prompt 661/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.650  max_d_mean=2.72e-03
+2026-07-20 13:45:15,532   prompt 662/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.789  max_d_mean=2.88e-03
+2026-07-20 13:45:20,185   prompt 663/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.560  max_d_mean=2.29e-03
+2026-07-20 13:45:24,684   prompt 664/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=2.129  max_d_mean=3.50e-03
+2026-07-20 13:45:29,229   prompt 665/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.687  max_d_mean=2.61e-03
+2026-07-20 13:45:33,792   prompt 666/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.012  max_d_mean=3.30e-03
+2026-07-20 13:45:38,313   prompt 667/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.985  max_d_mean=3.14e-03
+2026-07-20 13:45:42,876   prompt 668/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.374  max_d_mean=3.98e-03
+2026-07-20 13:45:47,884   prompt 669/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.109  max_d_mean=3.49e-03
+2026-07-20 13:45:52,843   prompt 670/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.982  max_d_mean=3.19e-03
+2026-07-20 13:45:57,934   prompt 671/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.642  max_d_mean=2.26e-03
+2026-07-20 13:46:02,574   prompt 672/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.585  max_d_mean=2.37e-03
+2026-07-20 13:46:07,083   prompt 673/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.856  max_d_mean=3.06e-03
+2026-07-20 13:46:11,805   prompt 674/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.729  max_d_mean=2.76e-03
+2026-07-20 13:46:16,391   prompt 675/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.551  max_d_mean=2.32e-03
+2026-07-20 13:46:21,115   prompt 676/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.501  max_d_mean=2.44e-03
+2026-07-20 13:46:25,635   prompt 677/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.005  max_d_mean=3.33e-03
+2026-07-20 13:46:30,374   prompt 678/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.128  max_d_mean=3.31e-03
+2026-07-20 13:46:34,976   prompt 679/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.702  max_d_mean=2.69e-03
+2026-07-20 13:46:39,694   prompt 680/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.710  max_d_mean=2.66e-03
+2026-07-20 13:46:44,315   prompt 681/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.753  max_d_mean=2.82e-03
+2026-07-20 13:46:48,822   prompt 682/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.687  max_d_mean=2.68e-03
+2026-07-20 13:46:53,689   prompt 683/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.573  max_d_mean=2.37e-03
+2026-07-20 13:46:58,189   prompt 684/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=2.007  max_d_mean=3.17e-03
+2026-07-20 13:47:02,852   prompt 685/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.623  max_d_mean=4.30e-03
+2026-07-20 13:47:07,356   prompt 686/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.855  max_d_mean=2.88e-03
+2026-07-20 13:47:11,964   prompt 687/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.608  max_d_mean=2.56e-03
+2026-07-20 13:47:16,639   prompt 688/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.799  max_d_mean=2.93e-03
+2026-07-20 13:47:21,296   prompt 689/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.022  max_d_mean=3.15e-03
+2026-07-20 13:47:25,976   prompt 690/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.619  max_d_mean=2.27e-03
+2026-07-20 13:47:30,786   prompt 691/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.929  max_d_mean=3.01e-03
+2026-07-20 13:47:35,371   prompt 692/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.917  max_d_mean=3.04e-03
+2026-07-20 13:47:39,985   prompt 693/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.287  max_d_mean=3.65e-03
+2026-07-20 13:47:44,654   prompt 694/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.340  max_d_mean=3.79e-03
+2026-07-20 13:47:49,279   prompt 695/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.027  max_d_mean=3.13e-03
+2026-07-20 13:47:54,118   prompt 696/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.372  max_d_mean=3.74e-03
+2026-07-20 13:47:58,642   prompt 697/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.024  max_d_mean=3.23e-03
+2026-07-20 13:48:03,371   prompt 698/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.852  max_d_mean=2.92e-03
+2026-07-20 13:48:08,066   prompt 699/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.718  max_d_mean=2.74e-03
+2026-07-20 13:48:12,799   prompt 700/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.719  max_d_mean=2.71e-03
+2026-07-20 13:48:17,596   prompt 701/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.836  max_d_mean=2.92e-03
+2026-07-20 13:48:22,256   prompt 702/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.044  max_d_mean=3.14e-03
+2026-07-20 13:48:26,902   prompt 703/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.798  max_d_mean=2.70e-03
+2026-07-20 13:48:31,701   prompt 704/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.498  max_d_mean=2.25e-03
+2026-07-20 13:48:36,273   prompt 705/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.787  max_d_mean=2.72e-03
+2026-07-20 13:48:41,025   prompt 706/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=3.743  max_d_mean=6.08e-03
+2026-07-20 13:48:45,683   prompt 707/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.178  max_d_mean=3.49e-03
+2026-07-20 13:48:50,269   prompt 708/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.584  max_d_mean=4.10e-03
+2026-07-20 13:48:55,021   prompt 709/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.897  max_d_mean=2.87e-03
+2026-07-20 13:49:00,007   prompt 710/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.050  max_d_mean=3.25e-03
+2026-07-20 13:49:04,592   prompt 711/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.058  max_d_mean=3.19e-03
+2026-07-20 13:49:09,275   prompt 712/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.487  max_d_mean=3.93e-03
+2026-07-20 13:49:13,793   prompt 713/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=3.159  max_d_mean=5.01e-03
+2026-07-20 13:49:18,401   prompt 714/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.376  max_d_mean=3.72e-03
+2026-07-20 13:49:23,015   prompt 715/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.634  max_d_mean=2.40e-03
+2026-07-20 13:49:27,779   prompt 716/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.607  max_d_mean=2.04e-03
+2026-07-20 13:49:32,335   prompt 717/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.459  max_d_mean=3.81e-03
+2026-07-20 13:49:37,392   prompt 718/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.835  max_d_mean=2.69e-03
+2026-07-20 13:49:42,063   prompt 719/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.062  max_d_mean=3.13e-03
+2026-07-20 13:49:46,602   prompt 720/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.042  max_d_mean=3.09e-03
+2026-07-20 13:49:51,128   prompt 721/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.132  max_d_mean=3.26e-03
+2026-07-20 13:49:55,704   prompt 722/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.948  max_d_mean=2.84e-03
+2026-07-20 13:50:00,312   prompt 723/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.926  max_d_mean=2.93e-03
+2026-07-20 13:50:04,911   prompt 724/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.266  max_d_mean=3.42e-03
+2026-07-20 13:50:09,493   prompt 725/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.814  max_d_mean=2.73e-03
+2026-07-20 13:50:14,487   prompt 726/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.692  max_d_mean=2.49e-03
+2026-07-20 13:50:19,298   prompt 727/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.795  max_d_mean=2.54e-03
+2026-07-20 13:50:23,968   prompt 728/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.126  max_d_mean=3.19e-03
+2026-07-20 13:50:28,687   prompt 729/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.019  max_d_mean=2.94e-03
+2026-07-20 13:50:33,190   prompt 730/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.849  max_d_mean=2.68e-03
+2026-07-20 13:50:37,727   prompt 731/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.930  max_d_mean=2.88e-03
+2026-07-20 13:50:42,291   prompt 732/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.318  max_d_mean=3.38e-03
+2026-07-20 13:50:46,971   prompt 733/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.960  max_d_mean=3.03e-03
+2026-07-20 13:50:51,534   prompt 734/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.268  max_d_mean=3.32e-03
+2026-07-20 13:50:56,134   prompt 735/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.897  max_d_mean=2.83e-03
+2026-07-20 13:51:00,984   prompt 736/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.856  max_d_mean=2.68e-03
+2026-07-20 13:51:05,795   prompt 737/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.489  max_d_mean=3.64e-03
+2026-07-20 13:51:10,586   prompt 738/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.134  max_d_mean=3.10e-03
+2026-07-20 13:51:15,100   prompt 739/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.598  max_d_mean=2.33e-03
+2026-07-20 13:51:19,672   prompt 740/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.007  max_d_mean=2.95e-03
+2026-07-20 13:51:24,280   prompt 741/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.780  max_d_mean=2.64e-03
+2026-07-20 13:51:28,909   prompt 742/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.678  max_d_mean=4.08e-03
+2026-07-20 13:51:33,523   prompt 743/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.108  max_d_mean=3.02e-03
+2026-07-20 13:51:38,172   prompt 744/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.805  max_d_mean=2.59e-03
+2026-07-20 13:51:42,681   prompt 745/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.752  max_d_mean=2.65e-03
+2026-07-20 13:51:47,401   prompt 746/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.426  max_d_mean=3.53e-03
+2026-07-20 13:51:52,028   prompt 747/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.769  max_d_mean=2.61e-03
+2026-07-20 13:51:56,791   prompt 748/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.340  max_d_mean=3.38e-03
+2026-07-20 13:52:01,587   prompt 749/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.014  max_d_mean=2.88e-03
+2026-07-20 13:52:06,209   prompt 750/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.096  max_d_mean=3.09e-03
+2026-07-20 13:52:10,834   prompt 751/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=2.090  max_d_mean=2.93e-03
+2026-07-20 13:52:15,682   prompt 752/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.269  max_d_mean=3.38e-03
+2026-07-20 13:52:20,235   prompt 753/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.990  max_d_mean=2.94e-03
+2026-07-20 13:52:24,893   prompt 754/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.806  max_d_mean=2.68e-03
+2026-07-20 13:52:29,396   prompt 755/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.020  max_d_mean=2.98e-03
+2026-07-20 13:52:34,286   prompt 756/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.642  max_d_mean=2.43e-03
+2026-07-20 13:52:38,923   prompt 757/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.779  max_d_mean=4.04e-03
+2026-07-20 13:52:43,603   prompt 758/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.302  max_d_mean=3.38e-03
+2026-07-20 13:52:48,387   prompt 759/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.327  max_d_mean=3.31e-03
+2026-07-20 13:52:53,045   prompt 760/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.708  max_d_mean=3.98e-03
+2026-07-20 13:52:57,652   prompt 761/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.432  max_d_mean=3.52e-03
+2026-07-20 13:53:02,315   prompt 762/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.376  max_d_mean=3.28e-03
+2026-07-20 13:53:06,886   prompt 763/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.729  max_d_mean=3.90e-03
+2026-07-20 13:53:11,645   prompt 764/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.236  max_d_mean=3.07e-03
+2026-07-20 13:53:16,312   prompt 765/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.216  max_d_mean=3.08e-03
+2026-07-20 13:53:20,896   prompt 766/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.782  max_d_mean=2.54e-03
+2026-07-20 13:53:25,607   prompt 767/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.217  max_d_mean=3.04e-03
+2026-07-20 13:53:30,336   prompt 768/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.102  max_d_mean=2.91e-03
+2026-07-20 13:53:35,027   prompt 769/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.164  max_d_mean=3.01e-03
+2026-07-20 13:53:39,544   prompt 770/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.366  max_d_mean=3.25e-03
+2026-07-20 13:53:44,231   prompt 771/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.290  max_d_mean=3.30e-03
+2026-07-20 13:53:48,901   prompt 772/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.341  max_d_mean=3.33e-03
+2026-07-20 13:53:53,410   prompt 773/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.221  max_d_mean=3.11e-03
+2026-07-20 13:53:58,096   prompt 774/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.761  max_d_mean=2.53e-03
+2026-07-20 13:54:02,698   prompt 775/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.878  max_d_mean=2.59e-03
+2026-07-20 13:54:07,502   prompt 776/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.142  max_d_mean=2.93e-03
+2026-07-20 13:54:12,525   prompt 777/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.131  max_d_mean=2.87e-03
+2026-07-20 13:54:17,184   prompt 778/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.599  max_d_mean=2.08e-03
+2026-07-20 13:54:21,895   prompt 779/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.590  max_d_mean=2.15e-03
+2026-07-20 13:54:26,379   prompt 780/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=1.721  max_d_mean=2.30e-03
+2026-07-20 13:54:31,374   prompt 781/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.727  max_d_mean=2.32e-03
+2026-07-20 13:54:36,184   prompt 782/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.720  max_d_mean=2.20e-03
+2026-07-20 13:54:40,812   prompt 783/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.594  max_d_mean=2.02e-03
+2026-07-20 13:54:45,499   prompt 784/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.882  max_d_mean=2.65e-03
+2026-07-20 13:54:50,052   prompt 785/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.939  max_d_mean=2.53e-03
+2026-07-20 13:54:54,770   prompt 786/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.644  max_d_mean=2.15e-03
+2026-07-20 13:54:59,364   prompt 787/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.816  max_d_mean=2.43e-03
+2026-07-20 13:55:03,982   prompt 788/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.092  max_d_mean=2.78e-03
+2026-07-20 13:55:08,649   prompt 789/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.121  max_d_mean=2.96e-03
+2026-07-20 13:55:13,215   prompt 790/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.948  max_d_mean=2.63e-03
+2026-07-20 13:55:17,702   prompt 791/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=2.130  max_d_mean=2.85e-03
+2026-07-20 13:55:22,305   prompt 792/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.926  max_d_mean=2.56e-03
+2026-07-20 13:55:26,899   prompt 793/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.841  max_d_mean=2.56e-03
+2026-07-20 13:55:31,593   prompt 794/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.904  max_d_mean=2.51e-03
+2026-07-20 13:55:36,139   prompt 795/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.993  max_d_mean=2.70e-03
+2026-07-20 13:55:40,991   prompt 796/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.053  max_d_mean=2.69e-03
+2026-07-20 13:55:45,487   prompt 797/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=1.932  max_d_mean=2.52e-03
+2026-07-20 13:55:50,105   prompt 798/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.883  max_d_mean=2.46e-03
+2026-07-20 13:55:54,600   prompt 799/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=1.649  max_d_mean=2.13e-03
+2026-07-20 13:55:59,276   prompt 800/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.166  max_d_mean=2.93e-03
+2026-07-20 13:56:03,995   prompt 801/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.195  max_d_mean=2.90e-03
+2026-07-20 13:56:08,725   prompt 802/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.710  max_d_mean=2.03e-03
+2026-07-20 13:56:13,616   prompt 803/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.960  max_d_mean=2.54e-03
+2026-07-20 13:56:18,191   prompt 804/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.184  max_d_mean=2.94e-03
+2026-07-20 13:56:22,899   prompt 805/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.635  max_d_mean=2.12e-03
+2026-07-20 13:56:27,549   prompt 806/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.304  max_d_mean=2.97e-03
+2026-07-20 13:56:32,281   prompt 807/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.705  max_d_mean=1.89e-03
+2026-07-20 13:56:36,787   prompt 808/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.790  max_d_mean=2.34e-03
+2026-07-20 13:56:41,480   prompt 809/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.713  max_d_mean=2.16e-03
+2026-07-20 13:56:46,109   prompt 810/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.766  max_d_mean=2.29e-03
+2026-07-20 13:56:50,792   prompt 811/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.608  max_d_mean=1.90e-03
+2026-07-20 13:56:55,293   prompt 812/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.535  max_d_mean=1.97e-03
+2026-07-20 13:56:59,987   prompt 813/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.689  max_d_mean=2.07e-03
+2026-07-20 13:57:04,524   prompt 814/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.236  max_d_mean=2.85e-03
+2026-07-20 13:57:09,277   prompt 815/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.723  max_d_mean=2.01e-03
+2026-07-20 13:57:13,898   prompt 816/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.687  max_d_mean=2.21e-03
+2026-07-20 13:57:18,550   prompt 817/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.080  max_d_mean=2.63e-03
+2026-07-20 13:57:23,135   prompt 818/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.910  max_d_mean=2.43e-03
+2026-07-20 13:57:27,829   prompt 819/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.922  max_d_mean=2.45e-03
+2026-07-20 13:57:32,694   prompt 820/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.864  max_d_mean=2.36e-03
+2026-07-20 13:57:37,421   prompt 821/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.846  max_d_mean=2.32e-03
+2026-07-20 13:57:42,221   prompt 822/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.173  max_d_mean=2.87e-03
+2026-07-20 13:57:47,391   prompt 823/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.228  max_d_mean=2.95e-03
+2026-07-20 13:57:52,087   prompt 824/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.554  max_d_mean=2.00e-03
+2026-07-20 13:57:56,695   prompt 825/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.614  max_d_mean=2.02e-03
+2026-07-20 13:58:01,479   prompt 826/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.797  max_d_mean=2.30e-03
+2026-07-20 13:58:06,121   prompt 827/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.662  max_d_mean=2.19e-03
+2026-07-20 13:58:10,796   prompt 828/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.973  max_d_mean=2.50e-03
+2026-07-20 13:58:15,708   prompt 829/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.185  max_d_mean=2.77e-03
+2026-07-20 13:58:20,653   prompt 830/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.565  max_d_mean=1.97e-03
+2026-07-20 13:58:25,149   prompt 831/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=1.796  max_d_mean=2.26e-03
+2026-07-20 13:58:29,781   prompt 832/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.127  max_d_mean=2.68e-03
+2026-07-20 13:58:34,400   prompt 833/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.349  max_d_mean=3.05e-03
+2026-07-20 13:58:38,902   prompt 834/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.617  max_d_mean=1.87e-03
+2026-07-20 13:58:43,644   prompt 835/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.917  max_d_mean=2.38e-03
+2026-07-20 13:58:48,567   prompt 836/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.905  max_d_mean=2.33e-03
+2026-07-20 13:58:53,513   prompt 837/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.576  max_d_mean=1.59e-03
+2026-07-20 13:58:57,990   prompt 838/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=1.561  max_d_mean=1.79e-03
+2026-07-20 13:59:02,501   prompt 839/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.810  max_d_mean=2.25e-03
+2026-07-20 13:59:07,093   prompt 840/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.593  max_d_mean=1.79e-03
+2026-07-20 13:59:11,677   prompt 841/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.284  max_d_mean=2.88e-03
+2026-07-20 13:59:16,279   prompt 842/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.015  max_d_mean=2.43e-03
+2026-07-20 13:59:21,509   prompt 843/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.620  max_d_mean=1.81e-03
+2026-07-20 13:59:26,000   prompt 844/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=1.611  max_d_mean=1.97e-03
+2026-07-20 13:59:30,520   prompt 845/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.540  max_d_mean=1.72e-03
+2026-07-20 13:59:35,088   prompt 846/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.606  max_d_mean=1.98e-03
+2026-07-20 13:59:39,683   prompt 847/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.605  max_d_mean=1.98e-03
+2026-07-20 13:59:44,364   prompt 848/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.668  max_d_mean=2.09e-03
+2026-07-20 13:59:49,011   prompt 849/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.685  max_d_mean=2.07e-03
+2026-07-20 13:59:53,510   prompt 850/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=1.606  max_d_mean=2.05e-03
+2026-07-20 13:59:58,179   prompt 851/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.991  max_d_mean=2.49e-03
+2026-07-20 14:00:02,690   prompt 852/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.645  max_d_mean=2.06e-03
+2026-07-20 14:00:07,217   prompt 853/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.684  max_d_mean=1.97e-03
+2026-07-20 14:00:12,031   prompt 854/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.552  max_d_mean=2.02e-03
+2026-07-20 14:00:16,718   prompt 855/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.658  max_d_mean=1.91e-03
+2026-07-20 14:00:21,393   prompt 856/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.576  max_d_mean=1.82e-03
+2026-07-20 14:00:25,927   prompt 857/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.633  max_d_mean=2.10e-03
+2026-07-20 14:00:30,500   prompt 858/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.284  max_d_mean=2.93e-03
+2026-07-20 14:00:35,075   prompt 859/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.731  max_d_mean=2.19e-03
+2026-07-20 14:00:39,624   prompt 860/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.013  max_d_mean=2.52e-03
+2026-07-20 14:00:44,315   prompt 861/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.850  max_d_mean=2.29e-03
+2026-07-20 14:00:49,195   prompt 862/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.780  max_d_mean=2.26e-03
+2026-07-20 14:00:53,984   prompt 863/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.619  max_d_mean=1.81e-03
+2026-07-20 14:00:58,479   prompt 864/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=2.089  max_d_mean=2.61e-03
+2026-07-20 14:01:02,982   prompt 865/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.693  max_d_mean=2.19e-03
+2026-07-20 14:01:07,619   prompt 866/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.889  max_d_mean=2.35e-03
+2026-07-20 14:01:12,198   prompt 867/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.898  max_d_mean=2.31e-03
+2026-07-20 14:01:16,864   prompt 868/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.262  max_d_mean=2.80e-03
+2026-07-20 14:01:21,541   prompt 869/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.063  max_d_mean=2.56e-03
+2026-07-20 14:01:26,098   prompt 870/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.817  max_d_mean=2.27e-03
+2026-07-20 14:01:30,693   prompt 871/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.207  max_d_mean=2.75e-03
+2026-07-20 14:01:35,336   prompt 872/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.614  max_d_mean=1.79e-03
+2026-07-20 14:01:39,889   prompt 873/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.899  max_d_mean=2.38e-03
+2026-07-20 14:01:44,997   prompt 874/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.069  max_d_mean=2.58e-03
+2026-07-20 14:01:49,709   prompt 875/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.876  max_d_mean=2.26e-03
+2026-07-20 14:01:54,600   prompt 876/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.905  max_d_mean=2.41e-03
+2026-07-20 14:01:59,179   prompt 877/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.194  max_d_mean=2.70e-03
+2026-07-20 14:02:03,747   prompt 878/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.720  max_d_mean=2.04e-03
+2026-07-20 14:02:08,388   prompt 879/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.598  max_d_mean=2.05e-03
+2026-07-20 14:02:13,328   prompt 880/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.782  max_d_mean=2.27e-03
+2026-07-20 14:02:18,109   prompt 881/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.846  max_d_mean=2.37e-03
+2026-07-20 14:02:22,581   prompt 882/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=1.565  max_d_mean=1.93e-03
+2026-07-20 14:02:27,448   prompt 883/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.643  max_d_mean=1.93e-03
+2026-07-20 14:02:31,965   prompt 884/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.909  max_d_mean=2.39e-03
+2026-07-20 14:02:36,611   prompt 885/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.994  max_d_mean=2.56e-03
+2026-07-20 14:02:41,276   prompt 886/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.775  max_d_mean=2.21e-03
+2026-07-20 14:02:45,788   prompt 887/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.940  max_d_mean=2.50e-03
+2026-07-20 14:02:50,278   prompt 888/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=1.959  max_d_mean=2.46e-03
+2026-07-20 14:02:54,969   prompt 889/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.777  max_d_mean=2.19e-03
+2026-07-20 14:02:59,622   prompt 890/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.212  max_d_mean=2.80e-03
+2026-07-20 14:03:04,297   prompt 891/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.813  max_d_mean=2.15e-03
+2026-07-20 14:03:08,892   prompt 892/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.488  max_d_mean=1.70e-03
+2026-07-20 14:03:13,505   prompt 893/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.769  max_d_mean=2.12e-03
+2026-07-20 14:03:18,290   prompt 894/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.719  max_d_mean=1.95e-03
+2026-07-20 14:03:22,778   prompt 895/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=1.896  max_d_mean=2.28e-03
+2026-07-20 14:03:27,396   prompt 896/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.575  max_d_mean=1.77e-03
+2026-07-20 14:03:32,199   prompt 897/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.829  max_d_mean=2.12e-03
+2026-07-20 14:03:36,903   prompt 898/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.609  max_d_mean=1.94e-03
+2026-07-20 14:03:41,375   prompt 899/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=1.927  max_d_mean=2.26e-03
+2026-07-20 14:03:46,079   prompt 900/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.156  max_d_mean=2.55e-03
+2026-07-20 14:03:50,895   prompt 901/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.099  max_d_mean=2.49e-03
+2026-07-20 14:03:55,392   prompt 902/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=1.568  max_d_mean=1.88e-03
+2026-07-20 14:04:00,267   prompt 903/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.114  max_d_mean=2.43e-03
+2026-07-20 14:04:04,826   prompt 904/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.852  max_d_mean=2.18e-03
+2026-07-20 14:04:09,365   prompt 905/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.627  max_d_mean=1.56e-03
+2026-07-20 14:04:13,904   prompt 906/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.632  max_d_mean=1.84e-03
+2026-07-20 14:04:18,570   prompt 907/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.765  max_d_mean=2.00e-03
+2026-07-20 14:04:23,078   prompt 908/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.037  max_d_mean=2.46e-03
+2026-07-20 14:04:27,680   prompt 909/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.977  max_d_mean=2.28e-03
+2026-07-20 14:04:32,175   prompt 910/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=1.736  max_d_mean=2.02e-03
+2026-07-20 14:04:36,686   prompt 911/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.860  max_d_mean=2.18e-03
+2026-07-20 14:04:41,419   prompt 912/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.743  max_d_mean=1.93e-03
+2026-07-20 14:04:46,099   prompt 913/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.676  max_d_mean=1.83e-03
+2026-07-20 14:04:50,897   prompt 914/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=3.038  max_d_mean=3.77e-03
+2026-07-20 14:04:55,410   prompt 915/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.155  max_d_mean=2.53e-03
+2026-07-20 14:05:00,035   prompt 916/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.073  max_d_mean=2.40e-03
+2026-07-20 14:05:04,798   prompt 917/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.522  max_d_mean=3.12e-03
+2026-07-20 14:05:09,606   prompt 918/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.498  max_d_mean=2.94e-03
+2026-07-20 14:05:14,190   prompt 919/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.908  max_d_mean=2.16e-03
+2026-07-20 14:05:18,794   prompt 920/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.876  max_d_mean=2.14e-03
+2026-07-20 14:05:23,386   prompt 921/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.535  max_d_mean=1.71e-03
+2026-07-20 14:05:27,975   prompt 922/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.787  max_d_mean=2.23e-03
+2026-07-20 14:05:32,900   prompt 923/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.819  max_d_mean=2.03e-03
+2026-07-20 14:05:37,710   prompt 924/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.878  max_d_mean=2.22e-03
+2026-07-20 14:05:42,296   prompt 925/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.774  max_d_mean=2.07e-03
+2026-07-20 14:05:47,275   prompt 926/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.693  max_d_mean=1.80e-03
+2026-07-20 14:05:51,975   prompt 927/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.788  max_d_mean=1.97e-03
+2026-07-20 14:05:56,553   prompt 928/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.898  max_d_mean=2.23e-03
+2026-07-20 14:06:01,201   prompt 929/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.984  max_d_mean=2.29e-03
+2026-07-20 14:06:05,800   prompt 930/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.026  max_d_mean=2.33e-03
+2026-07-20 14:06:10,385   prompt 931/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.804  max_d_mean=2.00e-03
+2026-07-20 14:06:14,900   prompt 932/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.961  max_d_mean=2.24e-03
+2026-07-20 14:06:19,468   prompt 933/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.816  max_d_mean=2.08e-03
+2026-07-20 14:06:24,090   prompt 934/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.198  max_d_mean=2.47e-03
+2026-07-20 14:06:28,905   prompt 935/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.801  max_d_mean=2.02e-03
+2026-07-20 14:06:33,597   prompt 936/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.663  max_d_mean=1.99e-03
+2026-07-20 14:06:38,184   prompt 937/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.552  max_d_mean=2.99e-03
+2026-07-20 14:06:42,780   prompt 938/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.585  max_d_mean=1.79e-03
+2026-07-20 14:06:47,449   prompt 939/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.591  max_d_mean=1.52e-03
+2026-07-20 14:06:52,028   prompt 940/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.744  max_d_mean=1.94e-03
+2026-07-20 14:06:56,689   prompt 941/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.735  max_d_mean=1.99e-03
+2026-07-20 14:07:01,489   prompt 942/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.011  max_d_mean=2.36e-03
+2026-07-20 14:07:06,405   prompt 943/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.167  max_d_mean=2.57e-03
+2026-07-20 14:07:11,170   prompt 944/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.766  max_d_mean=2.03e-03
+2026-07-20 14:07:15,733   prompt 945/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.916  max_d_mean=2.13e-03
+2026-07-20 14:07:20,413   prompt 946/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.618  max_d_mean=1.80e-03
+2026-07-20 14:07:25,031   prompt 947/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.568  max_d_mean=1.60e-03
+2026-07-20 14:07:29,544   prompt 948/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.981  max_d_mean=2.25e-03
+2026-07-20 14:07:34,169   prompt 949/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.557  max_d_mean=1.71e-03
+2026-07-20 14:07:38,677   prompt 950/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.848  max_d_mean=2.05e-03
+2026-07-20 14:07:43,413   prompt 951/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.679  max_d_mean=1.82e-03
+2026-07-20 14:07:47,983   prompt 952/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.530  max_d_mean=1.68e-03
+2026-07-20 14:07:52,491   prompt 953/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.931  max_d_mean=2.16e-03
+2026-07-20 14:07:57,193   prompt 954/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.644  max_d_mean=1.73e-03
+2026-07-20 14:08:01,890   prompt 955/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.683  max_d_mean=1.87e-03
+2026-07-20 14:08:06,467   prompt 956/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.591  max_d_mean=1.73e-03
+2026-07-20 14:08:11,182   prompt 957/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.562  max_d_mean=1.49e-03
+2026-07-20 14:08:15,712   prompt 958/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.697  max_d_mean=1.86e-03
+2026-07-20 14:08:20,293   prompt 959/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.615  max_d_mean=1.90e-03
+2026-07-20 14:08:25,202   prompt 960/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.791  max_d_mean=2.07e-03
+2026-07-20 14:08:29,788   prompt 961/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.515  max_d_mean=2.85e-03
+2026-07-20 14:08:34,315   prompt 962/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.621  max_d_mean=1.56e-03
+2026-07-20 14:08:38,875   prompt 963/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.294  max_d_mean=2.55e-03
+2026-07-20 14:08:43,445   prompt 964/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.617  max_d_mean=1.42e-03
+2026-07-20 14:08:48,096   prompt 965/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.115  max_d_mean=2.37e-03
+2026-07-20 14:08:52,603   prompt 966/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.881  max_d_mean=2.05e-03
+2026-07-20 14:08:57,235   prompt 967/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.557  max_d_mean=1.44e-03
+2026-07-20 14:09:02,093   prompt 968/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.646  max_d_mean=1.85e-03
+2026-07-20 14:09:06,575   prompt 969/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=1.926  max_d_mean=2.13e-03
+2026-07-20 14:09:11,080   prompt 970/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.847  max_d_mean=1.93e-03
+2026-07-20 14:09:15,579   prompt 971/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=1.731  max_d_mean=1.74e-03
+2026-07-20 14:09:20,083   prompt 972/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.450  max_d_mean=1.52e-03
+2026-07-20 14:09:24,920   prompt 973/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.463  max_d_mean=1.64e-03
+2026-07-20 14:09:29,483   prompt 974/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.634  max_d_mean=1.75e-03
+2026-07-20 14:09:33,996   prompt 975/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.578  max_d_mean=1.57e-03
+2026-07-20 14:09:38,708   prompt 976/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.012  max_d_mean=2.28e-03
+2026-07-20 14:09:43,384   prompt 977/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.682  max_d_mean=1.85e-03
+2026-07-20 14:09:47,885   prompt 978/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.609  max_d_mean=1.79e-03
+2026-07-20 14:09:52,407   prompt 979/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.653  max_d_mean=1.84e-03
+2026-07-20 14:09:57,087   prompt 980/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.560  max_d_mean=1.49e-03
+2026-07-20 14:10:01,691   prompt 981/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.628  max_d_mean=1.73e-03
+2026-07-20 14:10:06,188   prompt 982/1000  seq_len=128 n_valid=111  4s  max||J||/sqrt(d)=1.827  max_d_mean=2.00e-03
+2026-07-20 14:10:10,877   prompt 983/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.261  max_d_mean=2.44e-03
+2026-07-20 14:10:15,510   prompt 984/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.633  max_d_mean=1.51e-03
+2026-07-20 14:10:20,207   prompt 985/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.625  max_d_mean=1.55e-03
+2026-07-20 14:10:24,718   prompt 986/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.934  max_d_mean=2.09e-03
+2026-07-20 14:10:29,235   prompt 987/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.700  max_d_mean=1.80e-03
+2026-07-20 14:10:33,766   prompt 988/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.806  max_d_mean=1.99e-03
+2026-07-20 14:10:38,269   prompt 989/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.956  max_d_mean=2.05e-03
+2026-07-20 14:10:42,787   prompt 990/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.702  max_d_mean=1.62e-03
+2026-07-20 14:10:47,304   prompt 991/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.628  max_d_mean=1.80e-03
+2026-07-20 14:10:51,818   prompt 992/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.816  max_d_mean=1.98e-03
+2026-07-20 14:10:56,339   prompt 993/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.643  max_d_mean=1.47e-03
+2026-07-20 14:11:00,852   prompt 994/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.668  max_d_mean=1.45e-03
+2026-07-20 14:11:05,369   prompt 995/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.793  max_d_mean=1.84e-03
+2026-07-20 14:11:09,885   prompt 996/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.877  max_d_mean=2.04e-03
+2026-07-20 14:11:14,402   prompt 997/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.571  max_d_mean=1.60e-03
+2026-07-20 14:11:18,932   prompt 998/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.108  max_d_mean=2.30e-03
+2026-07-20 14:11:23,486   prompt 999/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=2.252  max_d_mean=2.39e-03
+2026-07-20 14:11:28,009   prompt 1000/1000  seq_len=128 n_valid=111  5s  max||J||/sqrt(d)=1.711  max_d_mean=1.84e-03
+2026-07-20 14:11:28,297 fit: done, 1000 prompts
+HFLensModel(LlamaForCausalLM, n_layers=16, d_model=2048)
+corpus: 1000 chunks (offset 0)
+fit done: 4643.0s total, 4.64s/prompt
+saved lenses/full_1b.pt