Add files using upload-large-folder tool

data/ulars/supplement/standard_eval/logs/LLM-LAT__zephyr7b-beta-rmu-lat-unlearn-wmdp-bio-cyber__wmdp_cyber.log

@@ -0,0 +1,22 @@
+repo_id: LLM-LAT/zephyr7b-beta-rmu-lat-unlearn-wmdp-bio-cyber
+split: cyber
+snapshot: /zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/LLM-LAT__zephyr7b-beta-rmu-lat-unlearn-wmdp-bio-cyber/models--LLM-LAT--zephyr7b-beta-rmu-lat-unlearn-wmdp-bio-cyber/snapshots/4ff066c804d0ab025698bb958c959c46642a1178
+cmd: /zfspool/zangyihe/conda_envs/corgo/bin/python src/eval.py experiment=eval/wmdp/default model=zephyr-7b-beta model.model_args.pretrained_model_name_or_path=/zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/LLM-LAT__zephyr7b-beta-rmu-lat-unlearn-wmdp-bio-cyber/models--LLM-LAT--zephyr7b-beta-rmu-lat-unlearn-wmdp-bio-cyber/snapshots/4ff066c804d0ab025698bb958c959c46642a1178 model.tokenizer_args.pretrained_model_name_or_path=/zfspool/zangyihe/ulars/models/zephyr-7b-beta model.model_args.attn_implementation=sdpa model.model_args.torch_dtype=bfloat16 data_split=cyber task_name=hf_wmdp_cyber_LLM-LAT__zephyr7b-beta-rmu-lat-unlearn-wmdp-bio-cyber paths.output_dir=/zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/LLM-LAT__zephyr7b-beta-rmu-lat-unlearn-wmdp-bio-cyber__wmdp_cyber eval.lm_eval.overwrite=true eval.lm_eval.tasks=[wmdp_cyber] eval.lm_eval.simple_evaluate_args.batch_size=16
+
+/zfspool/zangyihe/conda_envs/corgo/lib/python3.11/site-packages/transformers/utils/hub.py:105: FutureWarning: Using `TRANSFORMERS_CACHE` is deprecated and will be removed in v5 of Transformers. Use `HF_HOME` instead.
+  warnings.warn(
+[2026-05-30 00:10:50,333] [INFO] [real_accelerator.py:219:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+
+[2026-05-30 00:11:32,888][lm_eval.models.huggingface][WARNING] - `pretrained` model kwarg is not of type `str`. Many other model arguments may be ignored. Please do not launch via accelerate or use `parallelize=True` if passing an existing model this way.
+[2026-05-30 00:11:32,901][lm_eval.models.huggingface][WARNING] - Passed an already-initialized model through `pretrained`, assuming single-process call to evaluate() or custom distributed integration
+[2026-05-30 00:11:32,902][evaluator][INFO] - ***** Running LMEval evaluation suite *****
+[2026-05-30 00:11:32,902][evaluator][INFO] - Fine-grained evaluations will be saved to: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/LLM-LAT__zephyr7b-beta-rmu-lat-unlearn-wmdp-bio-cyber__wmdp_cyber/LMEval_EVAL.json
+[2026-05-30 00:11:32,902][evaluator][INFO] - Aggregated evaluations will be summarised in: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/LLM-LAT__zephyr7b-beta-rmu-lat-unlearn-wmdp-bio-cyber__wmdp_cyber/LMEval_SUMMARY.json
+[2026-05-30 00:11:32,903][lm_eval.evaluator][INFO] - Setting random seed to 0 | Setting numpy seed to 1234 | Setting torch manual seed to 1234 | Setting fewshot manual seed to 1234
+[2026-05-30 00:11:32,904][lm_eval.evaluator][INFO] - Using pre-initialized model
+[2026-05-30 00:11:38,123][lm_eval.tasks][INFO] - Selected tasks:
+[2026-05-30 00:11:38,123][lm_eval.tasks][INFO] - Task: wmdp_cyber (wmdp/wmdp_cyber.yaml)
+[2026-05-30 00:11:38,124][lm_eval.api.task][INFO] - Building contexts for wmdp_cyber on rank 0...
+
  0%|          | 0/1987 [00:00<?, ?it/s]
  5%|▍         | 96/1987 [00:00<00:01, 955.87it/s]
 10%|▉         | 195/1987 [00:00<00:01, 972.84it/s]
 15%|█▍        | 294/1987 [00:00<00:01, 980.19it/s]
 20%|█▉        | 393/1987 [00:00<00:01, 982.52it/s]
 25%|██▍       | 493/1987 [00:00<00:01, 987.40it/s]
 30%|██▉       | 593/1987 [00:00<00:01, 989.88it/s]
 35%|███▍      | 693/1987 [00:00<00:01, 991.76it/s]
 40%|███▉      | 794/1987 [00:00<00:01, 995.10it/s]
 45%|████▍     | 894/1987 [00:00<00:01, 995.19it/s]
 50%|█████     | 995/1987 [00:01<00:00, 997.20it/s]
 55%|█████▌    | 1096/1987 [00:01<00:00, 998.62it/s]
 60%|██████    | 1197/1987 [00:01<00:00, 999.52it/s]
 65%|██████▌   | 1298/1987 [00:01<00:00, 1000.77it/s]
 70%|███████   | 1399/1987 [00:01<00:00, 1001.97it/s]
 75%|███████▌  | 1500/1987 [00:01<00:00, 1002.16it/s]
 81%|████████  | 1601/1987 [00:01<00:00, 1002.08it/s]
 86%|████████▌ | 1702/1987 [00:01<00:00, 1001.01it/s]
 91%|█████████ | 1803/1987 [00:01<00:00, 999.45it/s] 
 96%|█████████▌| 1904/1987 [00:01<00:00, 999.95it/s]
+[2026-05-30 00:11:40,169][lm_eval.evaluator][INFO] - Running loglikelihood requests
+

data/ulars/supplement/standard_eval/logs/OPTML-Group__GradDiff-SAM-WMDP__wmdp_bio.log

@@ -0,0 +1,22 @@
+repo_id: OPTML-Group/GradDiff-SAM-WMDP
+split: bio
+snapshot: /zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/OPTML-Group__GradDiff-SAM-WMDP/models--OPTML-Group--GradDiff-SAM-WMDP/snapshots/e22b488a59005092ec1ec9e217ede68a183248be
+cmd: /zfspool/zangyihe/conda_envs/corgo/bin/python src/eval.py experiment=eval/wmdp/default model=zephyr-7b-beta model.model_args.pretrained_model_name_or_path=/zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/OPTML-Group__GradDiff-SAM-WMDP/models--OPTML-Group--GradDiff-SAM-WMDP/snapshots/e22b488a59005092ec1ec9e217ede68a183248be model.tokenizer_args.pretrained_model_name_or_path=/zfspool/zangyihe/ulars/models/zephyr-7b-beta model.model_args.attn_implementation=sdpa model.model_args.torch_dtype=bfloat16 data_split=bio task_name=hf_wmdp_bio_OPTML-Group__GradDiff-SAM-WMDP paths.output_dir=/zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/OPTML-Group__GradDiff-SAM-WMDP__wmdp_bio eval.lm_eval.overwrite=true eval.lm_eval.tasks=[wmdp_bio] eval.lm_eval.simple_evaluate_args.batch_size=16
+
+/zfspool/zangyihe/conda_envs/corgo/lib/python3.11/site-packages/transformers/utils/hub.py:105: FutureWarning: Using `TRANSFORMERS_CACHE` is deprecated and will be removed in v5 of Transformers. Use `HF_HOME` instead.
+  warnings.warn(
+[2026-05-29 21:52:59,303] [INFO] [real_accelerator.py:219:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+
+[2026-05-29 21:53:07,199][lm_eval.models.huggingface][WARNING] - `pretrained` model kwarg is not of type `str`. Many other model arguments may be ignored. Please do not launch via accelerate or use `parallelize=True` if passing an existing model this way.
+[2026-05-29 21:53:07,212][lm_eval.models.huggingface][WARNING] - Passed an already-initialized model through `pretrained`, assuming single-process call to evaluate() or custom distributed integration
+[2026-05-29 21:53:07,213][evaluator][INFO] - ***** Running LMEval evaluation suite *****
+[2026-05-29 21:53:07,213][evaluator][INFO] - Fine-grained evaluations will be saved to: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/OPTML-Group__GradDiff-SAM-WMDP__wmdp_bio/LMEval_EVAL.json
+[2026-05-29 21:53:07,213][evaluator][INFO] - Aggregated evaluations will be summarised in: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/OPTML-Group__GradDiff-SAM-WMDP__wmdp_bio/LMEval_SUMMARY.json
+[2026-05-29 21:53:07,214][lm_eval.evaluator][INFO] - Setting random seed to 0 | Setting numpy seed to 1234 | Setting torch manual seed to 1234 | Setting fewshot manual seed to 1234
+[2026-05-29 21:53:07,214][lm_eval.evaluator][INFO] - Using pre-initialized model
+[2026-05-29 21:53:28,578][lm_eval.tasks][INFO] - Selected tasks:
+[2026-05-29 21:53:28,578][lm_eval.tasks][INFO] - Task: wmdp_bio (wmdp/wmdp_bio.yaml)
+[2026-05-29 21:53:28,578][lm_eval.api.task][INFO] - Building contexts for wmdp_bio on rank 0...
+
  0%|          | 0/1273 [00:00<?, ?it/s]
  7%|▋         | 88/1273 [00:00<00:01, 876.68it/s]
 15%|█▍        | 186/1273 [00:00<00:01, 934.90it/s]
 22%|██▏       | 285/1273 [00:00<00:01, 958.38it/s]
 30%|███       | 384/1273 [00:00<00:00, 968.86it/s]
 38%|███▊      | 481/1273 [00:00<00:00, 962.07it/s]
 46%|████▌     | 581/1273 [00:00<00:00, 972.89it/s]
 53%|█████▎    | 679/1273 [00:00<00:00, 974.08it/s]
 61%|██████    | 779/1273 [00:00<00:00, 979.99it/s]
 69%|██████▉   | 879/1273 [00:00<00:00, 983.23it/s]
 77%|███████▋  | 979/1273 [00:01<00:00, 986.58it/s]
 85%|████████▍ | 1079/1273 [00:01<00:00, 989.23it/s]
 93%|█████████▎| 1179/1273 [00:01<00:00, 990.86it/s]
+[2026-05-29 21:53:29,912][lm_eval.evaluator][INFO] - Running loglikelihood requests
+

data/ulars/supplement/standard_eval/logs/OPTML-Group__GradDiff-SAM-WMDP__wmdp_cyber.log

@@ -0,0 +1,23 @@
+repo_id: OPTML-Group/GradDiff-SAM-WMDP
+split: cyber
+snapshot: /zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/OPTML-Group__GradDiff-SAM-WMDP/models--OPTML-Group--GradDiff-SAM-WMDP/snapshots/e22b488a59005092ec1ec9e217ede68a183248be
+cmd: /zfspool/zangyihe/conda_envs/corgo/bin/python src/eval.py experiment=eval/wmdp/default model=zephyr-7b-beta model.model_args.pretrained_model_name_or_path=/zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/OPTML-Group__GradDiff-SAM-WMDP/models--OPTML-Group--GradDiff-SAM-WMDP/snapshots/e22b488a59005092ec1ec9e217ede68a183248be model.tokenizer_args.pretrained_model_name_or_path=/zfspool/zangyihe/ulars/models/zephyr-7b-beta model.model_args.attn_implementation=sdpa model.model_args.torch_dtype=bfloat16 data_split=cyber task_name=hf_wmdp_cyber_OPTML-Group__GradDiff-SAM-WMDP paths.output_dir=/zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/OPTML-Group__GradDiff-SAM-WMDP__wmdp_cyber eval.lm_eval.overwrite=true eval.lm_eval.tasks=[wmdp_cyber] eval.lm_eval.simple_evaluate_args.batch_size=16
+
+/zfspool/zangyihe/conda_envs/corgo/lib/python3.11/site-packages/transformers/utils/hub.py:105: FutureWarning: Using `TRANSFORMERS_CACHE` is deprecated and will be removed in v5 of Transformers. Use `HF_HOME` instead.
+  warnings.warn(
+[2026-05-29 23:24:28,878] [INFO] [real_accelerator.py:219:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+
+[2026-05-29 23:24:52,802][lm_eval.models.huggingface][WARNING] - `pretrained` model kwarg is not of type `str`. Many other model arguments may be ignored. Please do not launch via accelerate or use `parallelize=True` if passing an existing model this way.
+[2026-05-29 23:24:52,814][lm_eval.models.huggingface][WARNING] - Passed an already-initialized model through `pretrained`, assuming single-process call to evaluate() or custom distributed integration
+[2026-05-29 23:24:52,815][evaluator][INFO] - ***** Running LMEval evaluation suite *****
+[2026-05-29 23:24:52,815][evaluator][INFO] - Fine-grained evaluations will be saved to: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/OPTML-Group__GradDiff-SAM-WMDP__wmdp_cyber/LMEval_EVAL.json
+[2026-05-29 23:24:52,815][evaluator][INFO] - Aggregated evaluations will be summarised in: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/OPTML-Group__GradDiff-SAM-WMDP__wmdp_cyber/LMEval_SUMMARY.json
+[2026-05-29 23:24:52,816][lm_eval.evaluator][INFO] - Setting random seed to 0 | Setting numpy seed to 1234 | Setting torch manual seed to 1234 | Setting fewshot manual seed to 1234
+[2026-05-29 23:24:52,816][lm_eval.evaluator][INFO] - Using pre-initialized model
+
+[2026-05-29 23:25:00,586][lm_eval.tasks][INFO] - Selected tasks:
+[2026-05-29 23:25:00,586][lm_eval.tasks][INFO] - Task: wmdp_cyber (wmdp/wmdp_cyber.yaml)
+[2026-05-29 23:25:00,587][lm_eval.api.task][INFO] - Building contexts for wmdp_cyber on rank 0...
+
  0%|          | 0/1987 [00:00<?, ?it/s]
  5%|▍         | 96/1987 [00:00<00:01, 956.68it/s]
 10%|▉         | 192/1987 [00:00<00:01, 955.81it/s]
 15%|█▍        | 292/1987 [00:00<00:01, 973.38it/s]
 20%|█▉        | 392/1987 [00:00<00:01, 981.69it/s]
 25%|██▍       | 492/1987 [00:00<00:01, 986.56it/s]
 30%|██▉       | 592/1987 [00:00<00:01, 990.09it/s]
 35%|███▍      | 692/1987 [00:00<00:01, 991.80it/s]
 40%|███▉      | 792/1987 [00:00<00:01, 992.07it/s]
 45%|████▍     | 892/1987 [00:00<00:01, 994.16it/s]
 50%|████▉     | 993/1987 [00:01<00:00, 998.66it/s]
 55%|█████▌    | 1094/1987 [00:01<00:00, 1000.54it/s]
 60%|██████    | 1195/1987 [00:01<00:00, 1002.17it/s]
 65%|██████▌   | 1296/1987 [00:01<00:00, 1001.85it/s]
 70%|███████   | 1397/1987 [00:01<00:00, 1003.95it/s]
 75%|███████▌  | 1498/1987 [00:01<00:00, 1004.54it/s]
 80%|████████  | 1599/1987 [00:01<00:00, 1003.26it/s]
 86%|████████▌ | 1700/1987 [00:01<00:00, 1004.07it/s]
 91%|█████████ | 1801/1987 [00:01<00:00, 1005.78it/s]
 96%|█████████▌| 1902/1987 [00:01<00:00, 1003.04it/s]
+[2026-05-29 23:25:02,631][lm_eval.evaluator][INFO] - Running loglikelihood requests
+

data/ulars/supplement/standard_eval/logs/OPTML-Group__GradDiff-WMDP__wmdp_bio.log

@@ -0,0 +1,22 @@
+repo_id: OPTML-Group/GradDiff-WMDP
+split: bio
+snapshot: /zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/OPTML-Group__GradDiff-WMDP/models--OPTML-Group--GradDiff-WMDP/snapshots/0695dccbb30a0986d20f211bc8bdea9d55ec5cc6
+cmd: /zfspool/zangyihe/conda_envs/corgo/bin/python src/eval.py experiment=eval/wmdp/default model=zephyr-7b-beta model.model_args.pretrained_model_name_or_path=/zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/OPTML-Group__GradDiff-WMDP/models--OPTML-Group--GradDiff-WMDP/snapshots/0695dccbb30a0986d20f211bc8bdea9d55ec5cc6 model.tokenizer_args.pretrained_model_name_or_path=/zfspool/zangyihe/ulars/models/zephyr-7b-beta model.model_args.attn_implementation=sdpa model.model_args.torch_dtype=bfloat16 data_split=bio task_name=hf_wmdp_bio_OPTML-Group__GradDiff-WMDP paths.output_dir=/zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/OPTML-Group__GradDiff-WMDP__wmdp_bio eval.lm_eval.overwrite=true eval.lm_eval.tasks=[wmdp_bio] eval.lm_eval.simple_evaluate_args.batch_size=16
+
+/zfspool/zangyihe/conda_envs/corgo/lib/python3.11/site-packages/transformers/utils/hub.py:105: FutureWarning: Using `TRANSFORMERS_CACHE` is deprecated and will be removed in v5 of Transformers. Use `HF_HOME` instead.
+  warnings.warn(
+[2026-05-29 21:52:58,725] [INFO] [real_accelerator.py:219:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+
+[2026-05-29 21:53:08,215][lm_eval.models.huggingface][WARNING] - `pretrained` model kwarg is not of type `str`. Many other model arguments may be ignored. Please do not launch via accelerate or use `parallelize=True` if passing an existing model this way.
+[2026-05-29 21:53:08,228][lm_eval.models.huggingface][WARNING] - Passed an already-initialized model through `pretrained`, assuming single-process call to evaluate() or custom distributed integration
+[2026-05-29 21:53:08,229][evaluator][INFO] - ***** Running LMEval evaluation suite *****
+[2026-05-29 21:53:08,229][evaluator][INFO] - Fine-grained evaluations will be saved to: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/OPTML-Group__GradDiff-WMDP__wmdp_bio/LMEval_EVAL.json
+[2026-05-29 21:53:08,229][evaluator][INFO] - Aggregated evaluations will be summarised in: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/OPTML-Group__GradDiff-WMDP__wmdp_bio/LMEval_SUMMARY.json
+[2026-05-29 21:53:08,231][lm_eval.evaluator][INFO] - Setting random seed to 0 | Setting numpy seed to 1234 | Setting torch manual seed to 1234 | Setting fewshot manual seed to 1234
+[2026-05-29 21:53:08,231][lm_eval.evaluator][INFO] - Using pre-initialized model
+[2026-05-29 21:53:16,388][lm_eval.tasks][INFO] - Selected tasks:
+[2026-05-29 21:53:16,388][lm_eval.tasks][INFO] - Task: wmdp_bio (wmdp/wmdp_bio.yaml)
+[2026-05-29 21:53:16,388][lm_eval.api.task][INFO] - Building contexts for wmdp_bio on rank 0...
+
  0%|          | 0/1273 [00:00<?, ?it/s]
  8%|▊         | 96/1273 [00:00<00:01, 955.73it/s]
 15%|█▌        | 195/1273 [00:00<00:01, 971.30it/s]
 23%|██▎       | 293/1273 [00:00<00:01, 959.01it/s]
 31%|███       | 392/1273 [00:00<00:00, 969.86it/s]
 38%|███▊      | 490/1273 [00:00<00:00, 967.35it/s]
 46%|████▋     | 590/1273 [00:00<00:00, 975.98it/s]
 54%|█████▍    | 690/1273 [00:00<00:00, 981.49it/s]
 62%|██████▏   | 790/1273 [00:00<00:00, 984.78it/s]
 70%|██████▉   | 889/1273 [00:00<00:00, 985.75it/s]
 78%|███████▊  | 988/1273 [00:01<00:00, 977.75it/s]
 85%|████████▌ | 1088/1273 [00:01<00:00, 982.36it/s]
 93%|█████████▎| 1188/1273 [00:01<00:00, 985.46it/s]
+[2026-05-29 21:53:17,720][lm_eval.evaluator][INFO] - Running loglikelihood requests
+

data/ulars/supplement/standard_eval/logs/OPTML-Group__GradDiff-WMDP__wmdp_cyber.log

@@ -0,0 +1,22 @@
+repo_id: OPTML-Group/GradDiff-WMDP
+split: cyber
+snapshot: /zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/OPTML-Group__GradDiff-WMDP/models--OPTML-Group--GradDiff-WMDP/snapshots/0695dccbb30a0986d20f211bc8bdea9d55ec5cc6
+cmd: /zfspool/zangyihe/conda_envs/corgo/bin/python src/eval.py experiment=eval/wmdp/default model=zephyr-7b-beta model.model_args.pretrained_model_name_or_path=/zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/OPTML-Group__GradDiff-WMDP/models--OPTML-Group--GradDiff-WMDP/snapshots/0695dccbb30a0986d20f211bc8bdea9d55ec5cc6 model.tokenizer_args.pretrained_model_name_or_path=/zfspool/zangyihe/ulars/models/zephyr-7b-beta model.model_args.attn_implementation=sdpa model.model_args.torch_dtype=bfloat16 data_split=cyber task_name=hf_wmdp_cyber_OPTML-Group__GradDiff-WMDP paths.output_dir=/zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/OPTML-Group__GradDiff-WMDP__wmdp_cyber eval.lm_eval.overwrite=true eval.lm_eval.tasks=[wmdp_cyber] eval.lm_eval.simple_evaluate_args.batch_size=16
+
+/zfspool/zangyihe/conda_envs/corgo/lib/python3.11/site-packages/transformers/utils/hub.py:105: FutureWarning: Using `TRANSFORMERS_CACHE` is deprecated and will be removed in v5 of Transformers. Use `HF_HOME` instead.
+  warnings.warn(
+[2026-05-29 23:27:04,280] [INFO] [real_accelerator.py:219:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+
+[2026-05-29 23:27:34,618][lm_eval.models.huggingface][WARNING] - `pretrained` model kwarg is not of type `str`. Many other model arguments may be ignored. Please do not launch via accelerate or use `parallelize=True` if passing an existing model this way.
+[2026-05-29 23:27:34,632][lm_eval.models.huggingface][WARNING] - Passed an already-initialized model through `pretrained`, assuming single-process call to evaluate() or custom distributed integration
+[2026-05-29 23:27:34,633][evaluator][INFO] - ***** Running LMEval evaluation suite *****
+[2026-05-29 23:27:34,633][evaluator][INFO] - Fine-grained evaluations will be saved to: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/OPTML-Group__GradDiff-WMDP__wmdp_cyber/LMEval_EVAL.json
+[2026-05-29 23:27:34,633][evaluator][INFO] - Aggregated evaluations will be summarised in: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/OPTML-Group__GradDiff-WMDP__wmdp_cyber/LMEval_SUMMARY.json
+[2026-05-29 23:27:34,634][lm_eval.evaluator][INFO] - Setting random seed to 0 | Setting numpy seed to 1234 | Setting torch manual seed to 1234 | Setting fewshot manual seed to 1234
+[2026-05-29 23:27:34,634][lm_eval.evaluator][INFO] - Using pre-initialized model
+[2026-05-29 23:27:40,222][lm_eval.tasks][INFO] - Selected tasks:
+[2026-05-29 23:27:40,222][lm_eval.tasks][INFO] - Task: wmdp_cyber (wmdp/wmdp_cyber.yaml)
+[2026-05-29 23:27:40,223][lm_eval.api.task][INFO] - Building contexts for wmdp_cyber on rank 0...
+
  0%|          | 0/1987 [00:00<?, ?it/s]
  5%|▍         | 96/1987 [00:00<00:01, 959.21it/s]
 10%|▉         | 194/1987 [00:00<00:01, 969.60it/s]
 15%|█▍        | 293/1987 [00:00<00:01, 978.18it/s]
 20%|█▉        | 393/1987 [00:00<00:01, 983.40it/s]
 25%|██▍       | 492/1987 [00:00<00:01, 979.67it/s]
 30%|██▉       | 592/1987 [00:00<00:01, 984.80it/s]
 35%|███▍      | 693/1987 [00:00<00:01, 990.17it/s]
 40%|███▉      | 793/1987 [00:00<00:01, 993.23it/s]
 45%|████▍     | 893/1987 [00:00<00:01, 992.86it/s]
 50%|████▉     | 993/1987 [00:01<00:01, 991.84it/s]
 55%|█████▌    | 1094/1987 [00:01<00:00, 994.86it/s]
 60%|██████    | 1195/1987 [00:01<00:00, 996.95it/s]
 65%|██████▌   | 1296/1987 [00:01<00:00, 999.51it/s]
 70%|███████   | 1397/1987 [00:01<00:00, 1000.82it/s]
 75%|███████▌  | 1498/1987 [00:01<00:00, 1000.86it/s]
 80%|████████  | 1599/1987 [00:01<00:00, 998.76it/s] 
 86%|████████▌ | 1699/1987 [00:01<00:00, 998.84it/s]
 91%|█████████ | 1799/1987 [00:01<00:00, 996.27it/s]
 96%|█████████▌| 1899/1987 [00:01<00:00, 996.17it/s]
+[2026-05-29 23:27:42,275][lm_eval.evaluator][INFO] - Running loglikelihood requests
+

data/ulars/supplement/standard_eval/logs/OPTML-Group__NPO-CR-WMDP__wmdp_bio.log

@@ -0,0 +1,22 @@
+repo_id: OPTML-Group/NPO-CR-WMDP
+split: bio
+snapshot: /zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/OPTML-Group__NPO-CR-WMDP/models--OPTML-Group--NPO-CR-WMDP/snapshots/731b22680ac91e7d24d6b10747b572e70825715a
+cmd: /zfspool/zangyihe/conda_envs/corgo/bin/python src/eval.py experiment=eval/wmdp/default model=zephyr-7b-beta model.model_args.pretrained_model_name_or_path=/zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/OPTML-Group__NPO-CR-WMDP/models--OPTML-Group--NPO-CR-WMDP/snapshots/731b22680ac91e7d24d6b10747b572e70825715a model.tokenizer_args.pretrained_model_name_or_path=/zfspool/zangyihe/ulars/models/zephyr-7b-beta model.model_args.attn_implementation=sdpa model.model_args.torch_dtype=bfloat16 data_split=bio task_name=hf_wmdp_bio_OPTML-Group__NPO-CR-WMDP paths.output_dir=/zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/OPTML-Group__NPO-CR-WMDP__wmdp_bio eval.lm_eval.overwrite=true eval.lm_eval.tasks=[wmdp_bio] eval.lm_eval.simple_evaluate_args.batch_size=16
+
+/zfspool/zangyihe/conda_envs/corgo/lib/python3.11/site-packages/transformers/utils/hub.py:105: FutureWarning: Using `TRANSFORMERS_CACHE` is deprecated and will be removed in v5 of Transformers. Use `HF_HOME` instead.
+  warnings.warn(
+[2026-05-29 21:52:58,890] [INFO] [real_accelerator.py:219:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+
+[2026-05-29 21:53:07,123][lm_eval.models.huggingface][WARNING] - `pretrained` model kwarg is not of type `str`. Many other model arguments may be ignored. Please do not launch via accelerate or use `parallelize=True` if passing an existing model this way.
+[2026-05-29 21:53:07,137][lm_eval.models.huggingface][WARNING] - Passed an already-initialized model through `pretrained`, assuming single-process call to evaluate() or custom distributed integration
+[2026-05-29 21:53:07,138][evaluator][INFO] - ***** Running LMEval evaluation suite *****
+[2026-05-29 21:53:07,138][evaluator][INFO] - Fine-grained evaluations will be saved to: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/OPTML-Group__NPO-CR-WMDP__wmdp_bio/LMEval_EVAL.json
+[2026-05-29 21:53:07,138][evaluator][INFO] - Aggregated evaluations will be summarised in: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/OPTML-Group__NPO-CR-WMDP__wmdp_bio/LMEval_SUMMARY.json
+[2026-05-29 21:53:07,140][lm_eval.evaluator][INFO] - Setting random seed to 0 | Setting numpy seed to 1234 | Setting torch manual seed to 1234 | Setting fewshot manual seed to 1234
+[2026-05-29 21:53:07,140][lm_eval.evaluator][INFO] - Using pre-initialized model
+[2026-05-29 21:53:16,390][lm_eval.tasks][INFO] - Selected tasks:
+[2026-05-29 21:53:16,390][lm_eval.tasks][INFO] - Task: wmdp_bio (wmdp/wmdp_bio.yaml)
+[2026-05-29 21:53:16,390][lm_eval.api.task][INFO] - Building contexts for wmdp_bio on rank 0...
+
  0%|          | 0/1273 [00:00<?, ?it/s]
  7%|▋         | 90/1273 [00:00<00:01, 897.30it/s]
 15%|█▍        | 188/1273 [00:00<00:01, 945.51it/s]
 23%|██▎       | 287/1273 [00:00<00:01, 964.16it/s]
 30%|███       | 384/1273 [00:00<00:00, 956.25it/s]
 38%|███▊      | 483/1273 [00:00<00:00, 967.02it/s]
 46%|████▌     | 581/1273 [00:00<00:00, 968.39it/s]
 53%|█████▎    | 680/1273 [00:00<00:00, 974.80it/s]
 61%|██████▏   | 780/1273 [00:00<00:00, 979.73it/s]
 69%|██████▉   | 878/1273 [00:00<00:00, 973.53it/s]
 77%|███████▋  | 977/1273 [00:01<00:00, 977.58it/s]
 85%|████████▍ | 1077/1273 [00:01<00:00, 982.47it/s]
 92%|█████████▏| 1176/1273 [00:01<00:00, 983.88it/s]
+[2026-05-29 21:53:17,729][lm_eval.evaluator][INFO] - Running loglikelihood requests
+

data/ulars/supplement/standard_eval/logs/OPTML-Group__NPO-CR-WMDP__wmdp_cyber.log

@@ -0,0 +1,22 @@
+repo_id: OPTML-Group/NPO-CR-WMDP
+split: cyber
+snapshot: /zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/OPTML-Group__NPO-CR-WMDP/models--OPTML-Group--NPO-CR-WMDP/snapshots/731b22680ac91e7d24d6b10747b572e70825715a
+cmd: /zfspool/zangyihe/conda_envs/corgo/bin/python src/eval.py experiment=eval/wmdp/default model=zephyr-7b-beta model.model_args.pretrained_model_name_or_path=/zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/OPTML-Group__NPO-CR-WMDP/models--OPTML-Group--NPO-CR-WMDP/snapshots/731b22680ac91e7d24d6b10747b572e70825715a model.tokenizer_args.pretrained_model_name_or_path=/zfspool/zangyihe/ulars/models/zephyr-7b-beta model.model_args.attn_implementation=sdpa model.model_args.torch_dtype=bfloat16 data_split=cyber task_name=hf_wmdp_cyber_OPTML-Group__NPO-CR-WMDP paths.output_dir=/zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/OPTML-Group__NPO-CR-WMDP__wmdp_cyber eval.lm_eval.overwrite=true eval.lm_eval.tasks=[wmdp_cyber] eval.lm_eval.simple_evaluate_args.batch_size=16
+
+/zfspool/zangyihe/conda_envs/corgo/lib/python3.11/site-packages/transformers/utils/hub.py:105: FutureWarning: Using `TRANSFORMERS_CACHE` is deprecated and will be removed in v5 of Transformers. Use `HF_HOME` instead.
+  warnings.warn(
+[2026-05-30 00:04:56,670] [INFO] [real_accelerator.py:219:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+
+[2026-05-30 00:05:41,998][lm_eval.models.huggingface][WARNING] - `pretrained` model kwarg is not of type `str`. Many other model arguments may be ignored. Please do not launch via accelerate or use `parallelize=True` if passing an existing model this way.
+[2026-05-30 00:05:42,011][lm_eval.models.huggingface][WARNING] - Passed an already-initialized model through `pretrained`, assuming single-process call to evaluate() or custom distributed integration
+[2026-05-30 00:05:42,012][evaluator][INFO] - ***** Running LMEval evaluation suite *****
+[2026-05-30 00:05:42,012][evaluator][INFO] - Fine-grained evaluations will be saved to: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/OPTML-Group__NPO-CR-WMDP__wmdp_cyber/LMEval_EVAL.json
+[2026-05-30 00:05:42,012][evaluator][INFO] - Aggregated evaluations will be summarised in: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/OPTML-Group__NPO-CR-WMDP__wmdp_cyber/LMEval_SUMMARY.json
+[2026-05-30 00:05:42,013][lm_eval.evaluator][INFO] - Setting random seed to 0 | Setting numpy seed to 1234 | Setting torch manual seed to 1234 | Setting fewshot manual seed to 1234
+[2026-05-30 00:05:42,014][lm_eval.evaluator][INFO] - Using pre-initialized model
+[2026-05-30 00:05:47,770][lm_eval.tasks][INFO] - Selected tasks:
+[2026-05-30 00:05:47,771][lm_eval.tasks][INFO] - Task: wmdp_cyber (wmdp/wmdp_cyber.yaml)
+[2026-05-30 00:05:47,771][lm_eval.api.task][INFO] - Building contexts for wmdp_cyber on rank 0...
+
  0%|          | 0/1987 [00:00<?, ?it/s]
  5%|▍         | 96/1987 [00:00<00:01, 951.81it/s]
 10%|▉         | 194/1987 [00:00<00:01, 967.05it/s]
 15%|█▍        | 293/1987 [00:00<00:01, 975.26it/s]
 20%|█▉        | 392/1987 [00:00<00:01, 978.37it/s]
 25%|██▍       | 491/1987 [00:00<00:01, 981.87it/s]
 30%|██▉       | 591/1987 [00:00<00:01, 985.16it/s]
 35%|███▍      | 690/1987 [00:00<00:01, 986.16it/s]
 40%|███▉      | 790/1987 [00:00<00:01, 988.53it/s]
 45%|████▍     | 890/1987 [00:00<00:01, 989.47it/s]
 50%|████▉     | 990/1987 [00:01<00:01, 989.91it/s]
 55%|█████▍    | 1090/1987 [00:01<00:00, 991.68it/s]
 60%|█████▉    | 1190/1987 [00:01<00:00, 991.45it/s]
 65%|██████▍   | 1290/1987 [00:01<00:00, 993.42it/s]
 70%|██████▉   | 1390/1987 [00:01<00:00, 993.89it/s]
 75%|███████▍  | 1490/1987 [00:01<00:00, 993.02it/s]
 80%|████████  | 1590/1987 [00:01<00:00, 993.42it/s]
 85%|████████▌ | 1690/1987 [00:01<00:00, 995.09it/s]
 90%|█████████ | 1790/1987 [00:01<00:00, 996.39it/s]
 95%|█████████▌| 1890/1987 [00:01<00:00, 996.78it/s]
+[2026-05-30 00:05:49,828][lm_eval.evaluator][INFO] - Running loglikelihood requests
+

data/ulars/supplement/standard_eval/logs/OPTML-Group__NPO-GP-WMDP__wmdp_bio.log

@@ -0,0 +1,22 @@
+repo_id: OPTML-Group/NPO-GP-WMDP
+split: bio
+snapshot: /zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/OPTML-Group__NPO-GP-WMDP/models--OPTML-Group--NPO-GP-WMDP/snapshots/9d1c75d774a7462edeb171bf0ea2ea3610fa67c5
+cmd: /zfspool/zangyihe/conda_envs/corgo/bin/python src/eval.py experiment=eval/wmdp/default model=zephyr-7b-beta model.model_args.pretrained_model_name_or_path=/zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/OPTML-Group__NPO-GP-WMDP/models--OPTML-Group--NPO-GP-WMDP/snapshots/9d1c75d774a7462edeb171bf0ea2ea3610fa67c5 model.tokenizer_args.pretrained_model_name_or_path=/zfspool/zangyihe/ulars/models/zephyr-7b-beta model.model_args.attn_implementation=sdpa model.model_args.torch_dtype=bfloat16 data_split=bio task_name=hf_wmdp_bio_OPTML-Group__NPO-GP-WMDP paths.output_dir=/zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/OPTML-Group__NPO-GP-WMDP__wmdp_bio eval.lm_eval.overwrite=true eval.lm_eval.tasks=[wmdp_bio] eval.lm_eval.simple_evaluate_args.batch_size=16
+
+/zfspool/zangyihe/conda_envs/corgo/lib/python3.11/site-packages/transformers/utils/hub.py:105: FutureWarning: Using `TRANSFORMERS_CACHE` is deprecated and will be removed in v5 of Transformers. Use `HF_HOME` instead.
+  warnings.warn(
+[2026-05-29 21:53:00,299] [INFO] [real_accelerator.py:219:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+
+[2026-05-29 21:53:08,205][lm_eval.models.huggingface][WARNING] - `pretrained` model kwarg is not of type `str`. Many other model arguments may be ignored. Please do not launch via accelerate or use `parallelize=True` if passing an existing model this way.
+[2026-05-29 21:53:08,218][lm_eval.models.huggingface][WARNING] - Passed an already-initialized model through `pretrained`, assuming single-process call to evaluate() or custom distributed integration
+[2026-05-29 21:53:08,219][evaluator][INFO] - ***** Running LMEval evaluation suite *****
+[2026-05-29 21:53:08,219][evaluator][INFO] - Fine-grained evaluations will be saved to: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/OPTML-Group__NPO-GP-WMDP__wmdp_bio/LMEval_EVAL.json
+[2026-05-29 21:53:08,220][evaluator][INFO] - Aggregated evaluations will be summarised in: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/OPTML-Group__NPO-GP-WMDP__wmdp_bio/LMEval_SUMMARY.json
+[2026-05-29 21:53:08,221][lm_eval.evaluator][INFO] - Setting random seed to 0 | Setting numpy seed to 1234 | Setting torch manual seed to 1234 | Setting fewshot manual seed to 1234
+[2026-05-29 21:53:08,221][lm_eval.evaluator][INFO] - Using pre-initialized model
+[2026-05-29 21:53:18,932][lm_eval.tasks][INFO] - Selected tasks:
+[2026-05-29 21:53:18,933][lm_eval.tasks][INFO] - Task: wmdp_bio (wmdp/wmdp_bio.yaml)
+[2026-05-29 21:53:18,933][lm_eval.api.task][INFO] - Building contexts for wmdp_bio on rank 0...
+
  0%|          | 0/1273 [00:00<?, ?it/s]
  8%|▊         | 96/1273 [00:00<00:01, 953.77it/s]
 15%|█▌        | 195/1273 [00:00<00:01, 973.56it/s]
 23%|██▎       | 295/1273 [00:00<00:00, 981.57it/s]
 31%|███       | 395/1273 [00:00<00:00, 988.07it/s]
 39%|███▉      | 495/1273 [00:00<00:00, 989.89it/s]
 47%|████▋     | 594/1273 [00:00<00:00, 980.21it/s]
 55%|█████▍    | 694/1273 [00:00<00:00, 985.48it/s]
 62%|██████▏   | 794/1273 [00:00<00:00, 989.98it/s]
 70%|███████   | 894/1273 [00:00<00:00, 991.45it/s]
 78%|███████▊  | 994/1273 [00:01<00:00, 993.25it/s]
 86%|████████▌ | 1094/1273 [00:01<00:00, 994.78it/s]
 94%|█████████▍| 1194/1273 [00:01<00:00, 996.11it/s]
+[2026-05-29 21:53:20,252][lm_eval.evaluator][INFO] - Running loglikelihood requests
+

data/ulars/supplement/standard_eval/logs/OPTML-Group__NPO-GP-WMDP__wmdp_cyber.log

@@ -0,0 +1,22 @@
+repo_id: OPTML-Group/NPO-GP-WMDP
+split: cyber
+snapshot: /zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/OPTML-Group__NPO-GP-WMDP/models--OPTML-Group--NPO-GP-WMDP/snapshots/9d1c75d774a7462edeb171bf0ea2ea3610fa67c5
+cmd: /zfspool/zangyihe/conda_envs/corgo/bin/python src/eval.py experiment=eval/wmdp/default model=zephyr-7b-beta model.model_args.pretrained_model_name_or_path=/zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/OPTML-Group__NPO-GP-WMDP/models--OPTML-Group--NPO-GP-WMDP/snapshots/9d1c75d774a7462edeb171bf0ea2ea3610fa67c5 model.tokenizer_args.pretrained_model_name_or_path=/zfspool/zangyihe/ulars/models/zephyr-7b-beta model.model_args.attn_implementation=sdpa model.model_args.torch_dtype=bfloat16 data_split=cyber task_name=hf_wmdp_cyber_OPTML-Group__NPO-GP-WMDP paths.output_dir=/zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/OPTML-Group__NPO-GP-WMDP__wmdp_cyber eval.lm_eval.overwrite=true eval.lm_eval.tasks=[wmdp_cyber] eval.lm_eval.simple_evaluate_args.batch_size=16
+
+/zfspool/zangyihe/conda_envs/corgo/lib/python3.11/site-packages/transformers/utils/hub.py:105: FutureWarning: Using `TRANSFORMERS_CACHE` is deprecated and will be removed in v5 of Transformers. Use `HF_HOME` instead.
+  warnings.warn(
+[2026-05-30 00:46:34,328] [INFO] [real_accelerator.py:219:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+
+[2026-05-30 00:47:03,219][lm_eval.models.huggingface][WARNING] - `pretrained` model kwarg is not of type `str`. Many other model arguments may be ignored. Please do not launch via accelerate or use `parallelize=True` if passing an existing model this way.
+[2026-05-30 00:47:03,232][lm_eval.models.huggingface][WARNING] - Passed an already-initialized model through `pretrained`, assuming single-process call to evaluate() or custom distributed integration
+[2026-05-30 00:47:03,233][evaluator][INFO] - ***** Running LMEval evaluation suite *****
+[2026-05-30 00:47:03,233][evaluator][INFO] - Fine-grained evaluations will be saved to: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/OPTML-Group__NPO-GP-WMDP__wmdp_cyber/LMEval_EVAL.json
+[2026-05-30 00:47:03,233][evaluator][INFO] - Aggregated evaluations will be summarised in: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/OPTML-Group__NPO-GP-WMDP__wmdp_cyber/LMEval_SUMMARY.json
+[2026-05-30 00:47:03,235][lm_eval.evaluator][INFO] - Setting random seed to 0 | Setting numpy seed to 1234 | Setting torch manual seed to 1234 | Setting fewshot manual seed to 1234
+[2026-05-30 00:47:03,235][lm_eval.evaluator][INFO] - Using pre-initialized model
+[2026-05-30 00:47:07,898][lm_eval.tasks][INFO] - Selected tasks:
+[2026-05-30 00:47:07,899][lm_eval.tasks][INFO] - Task: wmdp_cyber (wmdp/wmdp_cyber.yaml)
+[2026-05-30 00:47:07,899][lm_eval.api.task][INFO] - Building contexts for wmdp_cyber on rank 0...
+
  0%|          | 0/1987 [00:00<?, ?it/s]
  5%|▍         | 97/1987 [00:00<00:01, 965.18it/s]
 10%|▉         | 196/1987 [00:00<00:01, 978.82it/s]
 15%|█▍        | 296/1987 [00:00<00:01, 985.62it/s]
 20%|█▉        | 396/1987 [00:00<00:01, 989.18it/s]
 25%|██▌       | 497/1987 [00:00<00:01, 993.18it/s]
 30%|███       | 597/1987 [00:00<00:01, 995.13it/s]
 35%|███▌      | 698/1987 [00:00<00:01, 997.62it/s]
 40%|████      | 799/1987 [00:00<00:01, 1000.34it/s]
 45%|████▌     | 900/1987 [00:00<00:01, 992.22it/s] 
 50%|█████     | 1001/1987 [00:01<00:00, 995.25it/s]
 55%|█████▌    | 1102/1987 [00:01<00:00, 998.44it/s]
 61%|██████    | 1203/1987 [00:01<00:00, 1000.13it/s]
 66%|██████▌   | 1304/1987 [00:01<00:00, 1001.90it/s]
 71%|███████   | 1405/1987 [00:01<00:00, 1002.72it/s]
 76%|███████▌  | 1506/1987 [00:01<00:00, 1004.74it/s]
 81%|████████  | 1607/1987 [00:01<00:00, 1004.52it/s]
 86%|████████▌ | 1708/1987 [00:01<00:00, 998.63it/s] 
 91%|█████████ | 1809/1987 [00:01<00:00, 1000.64it/s]
 96%|█████████▌| 1910/1987 [00:01<00:00, 1001.72it/s]
+[2026-05-30 00:47:09,938][lm_eval.evaluator][INFO] - Running loglikelihood requests
+

data/ulars/supplement/standard_eval/logs/OPTML-Group__NPO-RS-WMDP__wmdp_bio.log

@@ -0,0 +1,22 @@
+repo_id: OPTML-Group/NPO-RS-WMDP
+split: bio
+snapshot: /zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/OPTML-Group__NPO-RS-WMDP/models--OPTML-Group--NPO-RS-WMDP/snapshots/beeeaa71af4dc91dcaf85333c4c3db340dcaab51
+cmd: /zfspool/zangyihe/conda_envs/corgo/bin/python src/eval.py experiment=eval/wmdp/default model=zephyr-7b-beta model.model_args.pretrained_model_name_or_path=/zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/OPTML-Group__NPO-RS-WMDP/models--OPTML-Group--NPO-RS-WMDP/snapshots/beeeaa71af4dc91dcaf85333c4c3db340dcaab51 model.tokenizer_args.pretrained_model_name_or_path=/zfspool/zangyihe/ulars/models/zephyr-7b-beta model.model_args.attn_implementation=sdpa model.model_args.torch_dtype=bfloat16 data_split=bio task_name=hf_wmdp_bio_OPTML-Group__NPO-RS-WMDP paths.output_dir=/zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/OPTML-Group__NPO-RS-WMDP__wmdp_bio eval.lm_eval.overwrite=true eval.lm_eval.tasks=[wmdp_bio] eval.lm_eval.simple_evaluate_args.batch_size=16
+
+/zfspool/zangyihe/conda_envs/corgo/lib/python3.11/site-packages/transformers/utils/hub.py:105: FutureWarning: Using `TRANSFORMERS_CACHE` is deprecated and will be removed in v5 of Transformers. Use `HF_HOME` instead.
+  warnings.warn(
+[2026-05-29 21:53:00,821] [INFO] [real_accelerator.py:219:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+
+[2026-05-29 21:53:08,914][lm_eval.models.huggingface][WARNING] - `pretrained` model kwarg is not of type `str`. Many other model arguments may be ignored. Please do not launch via accelerate or use `parallelize=True` if passing an existing model this way.
+[2026-05-29 21:53:08,926][lm_eval.models.huggingface][WARNING] - Passed an already-initialized model through `pretrained`, assuming single-process call to evaluate() or custom distributed integration
+[2026-05-29 21:53:08,927][evaluator][INFO] - ***** Running LMEval evaluation suite *****
+[2026-05-29 21:53:08,927][evaluator][INFO] - Fine-grained evaluations will be saved to: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/OPTML-Group__NPO-RS-WMDP__wmdp_bio/LMEval_EVAL.json
+[2026-05-29 21:53:08,927][evaluator][INFO] - Aggregated evaluations will be summarised in: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/OPTML-Group__NPO-RS-WMDP__wmdp_bio/LMEval_SUMMARY.json
+[2026-05-29 21:53:08,928][lm_eval.evaluator][INFO] - Setting random seed to 0 | Setting numpy seed to 1234 | Setting torch manual seed to 1234 | Setting fewshot manual seed to 1234
+[2026-05-29 21:53:08,928][lm_eval.evaluator][INFO] - Using pre-initialized model
+[2026-05-29 21:53:23,478][lm_eval.tasks][INFO] - Selected tasks:
+[2026-05-29 21:53:23,479][lm_eval.tasks][INFO] - Task: wmdp_bio (wmdp/wmdp_bio.yaml)
+[2026-05-29 21:53:23,479][lm_eval.api.task][INFO] - Building contexts for wmdp_bio on rank 0...
+
  0%|          | 0/1273 [00:00<?, ?it/s]
  8%|▊         | 96/1273 [00:00<00:01, 950.63it/s]
 15%|█▌        | 194/1273 [00:00<00:01, 964.84it/s]
 23%|██▎       | 291/1273 [00:00<00:01, 914.63it/s]
 31%|███       | 389/1273 [00:00<00:00, 938.30it/s]
 38%|███▊      | 488/1273 [00:00<00:00, 954.85it/s]
 46%|████▌     | 587/1273 [00:00<00:00, 965.10it/s]
 54%|█████▍    | 686/1273 [00:00<00:00, 971.70it/s]
 62%|██████▏   | 784/1273 [00:00<00:00, 964.91it/s]
 69%|██████▉   | 883/1273 [00:00<00:00, 971.99it/s]
 77%|███████▋  | 982/1273 [00:01<00:00, 976.75it/s]
 85%|████████▍ | 1080/1273 [00:01<00:00, 975.64it/s]
 93%|█████████▎| 1179/1273 [00:01<00:00, 979.73it/s]
+[2026-05-29 21:53:24,825][lm_eval.evaluator][INFO] - Running loglikelihood requests
+

data/ulars/supplement/standard_eval/logs/OPTML-Group__NPO-RS-WMDP__wmdp_cyber.log

@@ -0,0 +1,22 @@
+repo_id: OPTML-Group/NPO-RS-WMDP
+split: cyber
+snapshot: /zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/OPTML-Group__NPO-RS-WMDP/models--OPTML-Group--NPO-RS-WMDP/snapshots/beeeaa71af4dc91dcaf85333c4c3db340dcaab51
+cmd: /zfspool/zangyihe/conda_envs/corgo/bin/python src/eval.py experiment=eval/wmdp/default model=zephyr-7b-beta model.model_args.pretrained_model_name_or_path=/zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/OPTML-Group__NPO-RS-WMDP/models--OPTML-Group--NPO-RS-WMDP/snapshots/beeeaa71af4dc91dcaf85333c4c3db340dcaab51 model.tokenizer_args.pretrained_model_name_or_path=/zfspool/zangyihe/ulars/models/zephyr-7b-beta model.model_args.attn_implementation=sdpa model.model_args.torch_dtype=bfloat16 data_split=cyber task_name=hf_wmdp_cyber_OPTML-Group__NPO-RS-WMDP paths.output_dir=/zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/OPTML-Group__NPO-RS-WMDP__wmdp_cyber eval.lm_eval.overwrite=true eval.lm_eval.tasks=[wmdp_cyber] eval.lm_eval.simple_evaluate_args.batch_size=16
+
+/zfspool/zangyihe/conda_envs/corgo/lib/python3.11/site-packages/transformers/utils/hub.py:105: FutureWarning: Using `TRANSFORMERS_CACHE` is deprecated and will be removed in v5 of Transformers. Use `HF_HOME` instead.
+  warnings.warn(
+[2026-05-30 00:05:00,411] [INFO] [real_accelerator.py:219:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+
+[2026-05-30 00:05:39,314][lm_eval.models.huggingface][WARNING] - `pretrained` model kwarg is not of type `str`. Many other model arguments may be ignored. Please do not launch via accelerate or use `parallelize=True` if passing an existing model this way.
+[2026-05-30 00:05:39,327][lm_eval.models.huggingface][WARNING] - Passed an already-initialized model through `pretrained`, assuming single-process call to evaluate() or custom distributed integration
+[2026-05-30 00:05:39,328][evaluator][INFO] - ***** Running LMEval evaluation suite *****
+[2026-05-30 00:05:39,328][evaluator][INFO] - Fine-grained evaluations will be saved to: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/OPTML-Group__NPO-RS-WMDP__wmdp_cyber/LMEval_EVAL.json
+[2026-05-30 00:05:39,328][evaluator][INFO] - Aggregated evaluations will be summarised in: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/OPTML-Group__NPO-RS-WMDP__wmdp_cyber/LMEval_SUMMARY.json
+[2026-05-30 00:05:39,329][lm_eval.evaluator][INFO] - Setting random seed to 0 | Setting numpy seed to 1234 | Setting torch manual seed to 1234 | Setting fewshot manual seed to 1234
+[2026-05-30 00:05:39,329][lm_eval.evaluator][INFO] - Using pre-initialized model
+[2026-05-30 00:05:44,259][lm_eval.tasks][INFO] - Selected tasks:
+[2026-05-30 00:05:44,259][lm_eval.tasks][INFO] - Task: wmdp_cyber (wmdp/wmdp_cyber.yaml)
+[2026-05-30 00:05:44,260][lm_eval.api.task][INFO] - Building contexts for wmdp_cyber on rank 0...
+
  0%|          | 0/1987 [00:00<?, ?it/s]
  5%|▍         | 96/1987 [00:00<00:01, 955.26it/s]
 10%|▉         | 192/1987 [00:00<00:01, 943.56it/s]
 15%|█▍        | 291/1987 [00:00<00:01, 963.11it/s]
 20%|█▉        | 391/1987 [00:00<00:01, 973.75it/s]
 25%|██▍       | 491/1987 [00:00<00:01, 981.37it/s]
 30%|██▉       | 591/1987 [00:00<00:01, 985.74it/s]
 35%|███▍      | 691/1987 [00:00<00:01, 989.58it/s]
 40%|███▉      | 792/1987 [00:00<00:01, 993.99it/s]
 45%|████▍     | 892/1987 [00:00<00:01, 995.45it/s]
 50%|████▉     | 992/1987 [00:01<00:00, 996.38it/s]
 55%|█████▌    | 1093/1987 [00:01<00:00, 998.41it/s]
 60%|██████    | 1194/1987 [00:01<00:00, 999.21it/s]
 65%|██████▌   | 1295/1987 [00:01<00:00, 1001.40it/s]
 70%|███████   | 1396/1987 [00:01<00:00, 1000.82it/s]
 75%|███████▌  | 1497/1987 [00:01<00:00, 1001.36it/s]
 80%|████████  | 1598/1987 [00:01<00:00, 1000.91it/s]
 86%|████████▌ | 1699/1987 [00:01<00:00, 1001.62it/s]
 91%|█████████ | 1800/1987 [00:01<00:00, 1002.30it/s]
 96%|█████████▌| 1901/1987 [00:01<00:00, 1000.99it/s]
+[2026-05-30 00:05:46,310][lm_eval.evaluator][INFO] - Running loglikelihood requests
+

data/ulars/supplement/standard_eval/logs/OPTML-Group__NPO-SAM-WMDP__wmdp_bio.log

@@ -0,0 +1,23 @@
+repo_id: OPTML-Group/NPO-SAM-WMDP
+split: bio
+snapshot: /zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/OPTML-Group__NPO-SAM-WMDP/models--OPTML-Group--NPO-SAM-WMDP/snapshots/db68d7bef66ab8ac9c32b97ed960ee2ae2e48c25
+cmd: /zfspool/zangyihe/conda_envs/corgo/bin/python src/eval.py experiment=eval/wmdp/default model=zephyr-7b-beta model.model_args.pretrained_model_name_or_path=/zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/OPTML-Group__NPO-SAM-WMDP/models--OPTML-Group--NPO-SAM-WMDP/snapshots/db68d7bef66ab8ac9c32b97ed960ee2ae2e48c25 model.tokenizer_args.pretrained_model_name_or_path=/zfspool/zangyihe/ulars/models/zephyr-7b-beta model.model_args.attn_implementation=sdpa model.model_args.torch_dtype=bfloat16 data_split=bio task_name=hf_wmdp_bio_OPTML-Group__NPO-SAM-WMDP paths.output_dir=/zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/OPTML-Group__NPO-SAM-WMDP__wmdp_bio eval.lm_eval.overwrite=true eval.lm_eval.tasks=[wmdp_bio] eval.lm_eval.simple_evaluate_args.batch_size=16
+
+/zfspool/zangyihe/conda_envs/corgo/lib/python3.11/site-packages/transformers/utils/hub.py:105: FutureWarning: Using `TRANSFORMERS_CACHE` is deprecated and will be removed in v5 of Transformers. Use `HF_HOME` instead.
+  warnings.warn(
+[2026-05-29 21:52:59,840] [INFO] [real_accelerator.py:219:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+
+[2026-05-29 21:53:07,850][lm_eval.models.huggingface][WARNING] - `pretrained` model kwarg is not of type `str`. Many other model arguments may be ignored. Please do not launch via accelerate or use `parallelize=True` if passing an existing model this way.
+[2026-05-29 21:53:07,862][lm_eval.models.huggingface][WARNING] - Passed an already-initialized model through `pretrained`, assuming single-process call to evaluate() or custom distributed integration
+[2026-05-29 21:53:07,863][evaluator][INFO] - ***** Running LMEval evaluation suite *****
+[2026-05-29 21:53:07,863][evaluator][INFO] - Fine-grained evaluations will be saved to: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/OPTML-Group__NPO-SAM-WMDP__wmdp_bio/LMEval_EVAL.json
+[2026-05-29 21:53:07,863][evaluator][INFO] - Aggregated evaluations will be summarised in: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/OPTML-Group__NPO-SAM-WMDP__wmdp_bio/LMEval_SUMMARY.json
+[2026-05-29 21:53:07,864][lm_eval.evaluator][INFO] - Setting random seed to 0 | Setting numpy seed to 1234 | Setting torch manual seed to 1234 | Setting fewshot manual seed to 1234
+[2026-05-29 21:53:07,865][lm_eval.evaluator][INFO] - Using pre-initialized model
+
+[2026-05-29 21:53:16,340][lm_eval.tasks][INFO] - Selected tasks:
+[2026-05-29 21:53:16,341][lm_eval.tasks][INFO] - Task: wmdp_bio (wmdp/wmdp_bio.yaml)
+[2026-05-29 21:53:16,341][lm_eval.api.task][INFO] - Building contexts for wmdp_bio on rank 0...
+
  0%|          | 0/1273 [00:00<?, ?it/s]
  8%|▊         | 96/1273 [00:00<00:01, 952.82it/s]
 15%|█▌        | 195/1273 [00:00<00:01, 973.42it/s]
 23%|██▎       | 293/1273 [00:00<00:01, 973.91it/s]
 31%|███       | 393/1273 [00:00<00:00, 982.12it/s]
 39%|███▊      | 493/1273 [00:00<00:00, 987.32it/s]
 47%|████▋     | 593/1273 [00:00<00:00, 991.45it/s]
 55%|█████▍    | 694/1273 [00:00<00:00, 994.62it/s]
 62%|██████▏   | 794/1273 [00:00<00:00, 986.66it/s]
 70%|███████   | 894/1273 [00:00<00:00, 990.70it/s]
 78%|███████▊  | 994/1273 [00:01<00:00, 989.76it/s]
 86%|████████▌ | 1094/1273 [00:01<00:00, 992.75it/s]
 94%|█████████▍| 1195/1273 [00:01<00:00, 995.13it/s]
+[2026-05-29 21:53:17,659][lm_eval.evaluator][INFO] - Running loglikelihood requests
+

data/ulars/supplement/standard_eval/logs/OPTML-Group__NPO-SAM-WMDP__wmdp_cyber.log

@@ -0,0 +1,22 @@
+repo_id: OPTML-Group/NPO-SAM-WMDP
+split: cyber
+snapshot: /zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/OPTML-Group__NPO-SAM-WMDP/models--OPTML-Group--NPO-SAM-WMDP/snapshots/db68d7bef66ab8ac9c32b97ed960ee2ae2e48c25
+cmd: /zfspool/zangyihe/conda_envs/corgo/bin/python src/eval.py experiment=eval/wmdp/default model=zephyr-7b-beta model.model_args.pretrained_model_name_or_path=/zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/OPTML-Group__NPO-SAM-WMDP/models--OPTML-Group--NPO-SAM-WMDP/snapshots/db68d7bef66ab8ac9c32b97ed960ee2ae2e48c25 model.tokenizer_args.pretrained_model_name_or_path=/zfspool/zangyihe/ulars/models/zephyr-7b-beta model.model_args.attn_implementation=sdpa model.model_args.torch_dtype=bfloat16 data_split=cyber task_name=hf_wmdp_cyber_OPTML-Group__NPO-SAM-WMDP paths.output_dir=/zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/OPTML-Group__NPO-SAM-WMDP__wmdp_cyber eval.lm_eval.overwrite=true eval.lm_eval.tasks=[wmdp_cyber] eval.lm_eval.simple_evaluate_args.batch_size=16
+
+/zfspool/zangyihe/conda_envs/corgo/lib/python3.11/site-packages/transformers/utils/hub.py:105: FutureWarning: Using `TRANSFORMERS_CACHE` is deprecated and will be removed in v5 of Transformers. Use `HF_HOME` instead.
+  warnings.warn(
+[2026-05-30 00:07:49,498] [INFO] [real_accelerator.py:219:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+
+[2026-05-30 00:08:17,087][lm_eval.models.huggingface][WARNING] - `pretrained` model kwarg is not of type `str`. Many other model arguments may be ignored. Please do not launch via accelerate or use `parallelize=True` if passing an existing model this way.
+[2026-05-30 00:08:17,104][lm_eval.models.huggingface][WARNING] - Passed an already-initialized model through `pretrained`, assuming single-process call to evaluate() or custom distributed integration
+[2026-05-30 00:08:17,105][evaluator][INFO] - ***** Running LMEval evaluation suite *****
+[2026-05-30 00:08:17,105][evaluator][INFO] - Fine-grained evaluations will be saved to: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/OPTML-Group__NPO-SAM-WMDP__wmdp_cyber/LMEval_EVAL.json
+[2026-05-30 00:08:17,105][evaluator][INFO] - Aggregated evaluations will be summarised in: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/OPTML-Group__NPO-SAM-WMDP__wmdp_cyber/LMEval_SUMMARY.json
+[2026-05-30 00:08:17,106][lm_eval.evaluator][INFO] - Setting random seed to 0 | Setting numpy seed to 1234 | Setting torch manual seed to 1234 | Setting fewshot manual seed to 1234
+[2026-05-30 00:08:17,106][lm_eval.evaluator][INFO] - Using pre-initialized model
+[2026-05-30 00:08:25,195][lm_eval.tasks][INFO] - Selected tasks:
+[2026-05-30 00:08:25,196][lm_eval.tasks][INFO] - Task: wmdp_cyber (wmdp/wmdp_cyber.yaml)
+[2026-05-30 00:08:25,196][lm_eval.api.task][INFO] - Building contexts for wmdp_cyber on rank 0...
+
  0%|          | 0/1987 [00:00<?, ?it/s]
  5%|▍         | 94/1987 [00:00<00:02, 936.99it/s]
  9%|▉         | 188/1987 [00:00<00:01, 938.49it/s]
 14%|█▍        | 285/1987 [00:00<00:01, 949.42it/s]
 19%|█▉        | 381/1987 [00:00<00:01, 953.50it/s]
 24%|██▍       | 478/1987 [00:00<00:01, 958.88it/s]
 29%|██▉       | 576/1987 [00:00<00:01, 965.35it/s]
 34%|███▍      | 673/1987 [00:00<00:01, 966.80it/s]
 39%|███▉      | 772/1987 [00:00<00:01, 972.18it/s]
 44%|████▍     | 871/1987 [00:00<00:01, 975.84it/s]
 49%|████▉     | 970/1987 [00:01<00:01, 978.79it/s]
 54%|█████▍    | 1069/1987 [00:01<00:00, 981.72it/s]
 59%|█████▉    | 1168/1987 [00:01<00:00, 982.11it/s]
 64%|██████▍   | 1267/1987 [00:01<00:00, 984.45it/s]
 69%|██████▉   | 1367/1987 [00:01<00:00, 986.60it/s]
 74%|███████▍  | 1466/1987 [00:01<00:00, 985.01it/s]
 79%|███████▉  | 1566/1987 [00:01<00:00, 987.53it/s]
 84%|████████▍ | 1665/1987 [00:01<00:00, 986.93it/s]
 89%|████████▉ | 1764/1987 [00:01<00:00, 987.37it/s]
 94%|█████████▍| 1863/1987 [00:01<00:00, 987.25it/s]
 99%|█████████▊| 1962/1987 [00:02<00:00, 981.04it/s]
+[2026-05-30 00:08:27,285][lm_eval.evaluator][INFO] - Running loglikelihood requests
+

data/ulars/supplement/standard_eval/logs/OPTML-Group__NPO-WA-WMDP__wmdp_bio.log

@@ -0,0 +1,22 @@
+repo_id: OPTML-Group/NPO-WA-WMDP
+split: bio
+snapshot: /zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/OPTML-Group__NPO-WA-WMDP/models--OPTML-Group--NPO-WA-WMDP/snapshots/6edffa2b4f929608aa3ff876b5853363c9f75d8b
+cmd: /zfspool/zangyihe/conda_envs/corgo/bin/python src/eval.py experiment=eval/wmdp/default model=zephyr-7b-beta model.model_args.pretrained_model_name_or_path=/zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/OPTML-Group__NPO-WA-WMDP/models--OPTML-Group--NPO-WA-WMDP/snapshots/6edffa2b4f929608aa3ff876b5853363c9f75d8b model.tokenizer_args.pretrained_model_name_or_path=/zfspool/zangyihe/ulars/models/zephyr-7b-beta model.model_args.attn_implementation=sdpa model.model_args.torch_dtype=bfloat16 data_split=bio task_name=hf_wmdp_bio_OPTML-Group__NPO-WA-WMDP paths.output_dir=/zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/OPTML-Group__NPO-WA-WMDP__wmdp_bio eval.lm_eval.overwrite=true eval.lm_eval.tasks=[wmdp_bio] eval.lm_eval.simple_evaluate_args.batch_size=16
+
+/zfspool/zangyihe/conda_envs/corgo/lib/python3.11/site-packages/transformers/utils/hub.py:105: FutureWarning: Using `TRANSFORMERS_CACHE` is deprecated and will be removed in v5 of Transformers. Use `HF_HOME` instead.
+  warnings.warn(
+[2026-05-29 21:57:30,850] [INFO] [real_accelerator.py:219:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+
+[2026-05-29 21:57:55,426][lm_eval.models.huggingface][WARNING] - `pretrained` model kwarg is not of type `str`. Many other model arguments may be ignored. Please do not launch via accelerate or use `parallelize=True` if passing an existing model this way.
+[2026-05-29 21:57:55,438][lm_eval.models.huggingface][WARNING] - Passed an already-initialized model through `pretrained`, assuming single-process call to evaluate() or custom distributed integration
+[2026-05-29 21:57:55,439][evaluator][INFO] - ***** Running LMEval evaluation suite *****
+[2026-05-29 21:57:55,439][evaluator][INFO] - Fine-grained evaluations will be saved to: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/OPTML-Group__NPO-WA-WMDP__wmdp_bio/LMEval_EVAL.json
+[2026-05-29 21:57:55,439][evaluator][INFO] - Aggregated evaluations will be summarised in: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/OPTML-Group__NPO-WA-WMDP__wmdp_bio/LMEval_SUMMARY.json
+[2026-05-29 21:57:55,441][lm_eval.evaluator][INFO] - Setting random seed to 0 | Setting numpy seed to 1234 | Setting torch manual seed to 1234 | Setting fewshot manual seed to 1234
+[2026-05-29 21:57:55,441][lm_eval.evaluator][INFO] - Using pre-initialized model
+[2026-05-29 21:58:09,547][lm_eval.tasks][INFO] - Selected tasks:
+[2026-05-29 21:58:09,548][lm_eval.tasks][INFO] - Task: wmdp_bio (wmdp/wmdp_bio.yaml)
+[2026-05-29 21:58:09,548][lm_eval.api.task][INFO] - Building contexts for wmdp_bio on rank 0...
+
  0%|          | 0/1273 [00:00<?, ?it/s]
  7%|▋         | 95/1273 [00:00<00:01, 944.55it/s]
 15%|█▌        | 192/1273 [00:00<00:01, 957.31it/s]
 23%|██▎       | 290/1273 [00:00<00:01, 964.75it/s]
 30%|███       | 387/1273 [00:00<00:00, 958.55it/s]
 38%|███▊      | 485/1273 [00:00<00:00, 964.97it/s]
 46%|████▌     | 584/1273 [00:00<00:00, 973.38it/s]
 54%|█████▎    | 682/1273 [00:00<00:00, 974.72it/s]
 61%|██████▏   | 781/1273 [00:00<00:00, 977.56it/s]
 69%|██████▉   | 879/1273 [00:00<00:00, 972.57it/s]
 77%|███████▋  | 978/1273 [00:01<00:00, 975.88it/s]
 85%|████████▍ | 1077/1273 [00:01<00:00, 978.09it/s]
 92%|█████████▏| 1175/1273 [00:01<00:00, 974.14it/s]
+[2026-05-29 21:58:10,888][lm_eval.evaluator][INFO] - Running loglikelihood requests
+

data/ulars/supplement/standard_eval/logs/OPTML-Group__NPO-WA-WMDP__wmdp_cyber.log

@@ -0,0 +1,22 @@
+repo_id: OPTML-Group/NPO-WA-WMDP
+split: cyber
+snapshot: /zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/OPTML-Group__NPO-WA-WMDP/models--OPTML-Group--NPO-WA-WMDP/snapshots/6edffa2b4f929608aa3ff876b5853363c9f75d8b
+cmd: /zfspool/zangyihe/conda_envs/corgo/bin/python src/eval.py experiment=eval/wmdp/default model=zephyr-7b-beta model.model_args.pretrained_model_name_or_path=/zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/OPTML-Group__NPO-WA-WMDP/models--OPTML-Group--NPO-WA-WMDP/snapshots/6edffa2b4f929608aa3ff876b5853363c9f75d8b model.tokenizer_args.pretrained_model_name_or_path=/zfspool/zangyihe/ulars/models/zephyr-7b-beta model.model_args.attn_implementation=sdpa model.model_args.torch_dtype=bfloat16 data_split=cyber task_name=hf_wmdp_cyber_OPTML-Group__NPO-WA-WMDP paths.output_dir=/zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/OPTML-Group__NPO-WA-WMDP__wmdp_cyber eval.lm_eval.overwrite=true eval.lm_eval.tasks=[wmdp_cyber] eval.lm_eval.simple_evaluate_args.batch_size=16
+
+/zfspool/zangyihe/conda_envs/corgo/lib/python3.11/site-packages/transformers/utils/hub.py:105: FutureWarning: Using `TRANSFORMERS_CACHE` is deprecated and will be removed in v5 of Transformers. Use `HF_HOME` instead.
+  warnings.warn(
+[2026-05-30 00:07:51,056] [INFO] [real_accelerator.py:219:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+
+[2026-05-30 00:08:35,378][lm_eval.models.huggingface][WARNING] - `pretrained` model kwarg is not of type `str`. Many other model arguments may be ignored. Please do not launch via accelerate or use `parallelize=True` if passing an existing model this way.
+[2026-05-30 00:08:35,393][lm_eval.models.huggingface][WARNING] - Passed an already-initialized model through `pretrained`, assuming single-process call to evaluate() or custom distributed integration
+[2026-05-30 00:08:35,394][evaluator][INFO] - ***** Running LMEval evaluation suite *****
+[2026-05-30 00:08:35,395][evaluator][INFO] - Fine-grained evaluations will be saved to: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/OPTML-Group__NPO-WA-WMDP__wmdp_cyber/LMEval_EVAL.json
+[2026-05-30 00:08:35,395][evaluator][INFO] - Aggregated evaluations will be summarised in: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/OPTML-Group__NPO-WA-WMDP__wmdp_cyber/LMEval_SUMMARY.json
+[2026-05-30 00:08:35,396][lm_eval.evaluator][INFO] - Setting random seed to 0 | Setting numpy seed to 1234 | Setting torch manual seed to 1234 | Setting fewshot manual seed to 1234
+[2026-05-30 00:08:35,396][lm_eval.evaluator][INFO] - Using pre-initialized model
+[2026-05-30 00:08:47,525][lm_eval.tasks][INFO] - Selected tasks:
+[2026-05-30 00:08:47,525][lm_eval.tasks][INFO] - Task: wmdp_cyber (wmdp/wmdp_cyber.yaml)
+[2026-05-30 00:08:47,525][lm_eval.api.task][INFO] - Building contexts for wmdp_cyber on rank 0...
+
  0%|          | 0/1987 [00:00<?, ?it/s]
  5%|▍         | 96/1987 [00:00<00:01, 954.33it/s]
 10%|▉         | 194/1987 [00:00<00:01, 965.27it/s]
 15%|█▍        | 291/1987 [00:00<00:01, 961.24it/s]
 20%|█▉        | 390/1987 [00:00<00:01, 969.16it/s]
 25%|██▍       | 487/1987 [00:00<00:01, 968.13it/s]
 29%|██▉       | 585/1987 [00:00<00:01, 971.47it/s]
 34%|███▍      | 683/1987 [00:00<00:01, 951.96it/s]
 39%|███▉      | 781/1987 [00:00<00:01, 960.11it/s]
 44%|████▍     | 880/1987 [00:00<00:01, 967.58it/s]
 49%|████▉     | 979/1987 [00:01<00:01, 973.50it/s]
 54%|█████▍    | 1078/1987 [00:01<00:00, 978.03it/s]
 59%|█████▉    | 1177/1987 [00:01<00:00, 981.17it/s]
 64%|██████▍   | 1277/1987 [00:01<00:00, 983.92it/s]
 69%|██████▉   | 1377/1987 [00:01<00:00, 985.88it/s]
 74%|███████▍  | 1477/1987 [00:01<00:00, 988.32it/s]
 79%|███████▉  | 1577/1987 [00:01<00:00, 989.84it/s]
 84%|████████▍ | 1677/1987 [00:01<00:00, 990.75it/s]
 89%|████████▉ | 1777/1987 [00:01<00:00, 991.34it/s]
 94%|█████████▍| 1877/1987 [00:01<00:00, 991.93it/s]
 99%|█████████▉| 1977/1987 [00:02<00:00, 990.17it/s]
+[2026-05-30 00:08:49,604][lm_eval.evaluator][INFO] - Running loglikelihood requests
+

data/ulars/supplement/standard_eval/logs/OPTML-Group__NPO-WMDP__wmdp_bio.log

@@ -0,0 +1,22 @@
+repo_id: OPTML-Group/NPO-WMDP
+split: bio
+snapshot: /zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/OPTML-Group__NPO-WMDP/models--OPTML-Group--NPO-WMDP/snapshots/fec7960fb3925037a6befc5819036cbc55664b18
+cmd: /zfspool/zangyihe/conda_envs/corgo/bin/python src/eval.py experiment=eval/wmdp/default model=zephyr-7b-beta model.model_args.pretrained_model_name_or_path=/zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/OPTML-Group__NPO-WMDP/models--OPTML-Group--NPO-WMDP/snapshots/fec7960fb3925037a6befc5819036cbc55664b18 model.tokenizer_args.pretrained_model_name_or_path=/zfspool/zangyihe/ulars/models/zephyr-7b-beta model.model_args.attn_implementation=sdpa model.model_args.torch_dtype=bfloat16 data_split=bio task_name=hf_wmdp_bio_OPTML-Group__NPO-WMDP paths.output_dir=/zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/OPTML-Group__NPO-WMDP__wmdp_bio eval.lm_eval.overwrite=true eval.lm_eval.tasks=[wmdp_bio] eval.lm_eval.simple_evaluate_args.batch_size=16
+
+/zfspool/zangyihe/conda_envs/corgo/lib/python3.11/site-packages/transformers/utils/hub.py:105: FutureWarning: Using `TRANSFORMERS_CACHE` is deprecated and will be removed in v5 of Transformers. Use `HF_HOME` instead.
+  warnings.warn(
+[2026-05-29 21:57:55,071] [INFO] [real_accelerator.py:219:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+
+[2026-05-29 21:58:28,699][lm_eval.models.huggingface][WARNING] - `pretrained` model kwarg is not of type `str`. Many other model arguments may be ignored. Please do not launch via accelerate or use `parallelize=True` if passing an existing model this way.
+[2026-05-29 21:58:28,711][lm_eval.models.huggingface][WARNING] - Passed an already-initialized model through `pretrained`, assuming single-process call to evaluate() or custom distributed integration
+[2026-05-29 21:58:28,712][evaluator][INFO] - ***** Running LMEval evaluation suite *****
+[2026-05-29 21:58:28,712][evaluator][INFO] - Fine-grained evaluations will be saved to: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/OPTML-Group__NPO-WMDP__wmdp_bio/LMEval_EVAL.json
+[2026-05-29 21:58:28,712][evaluator][INFO] - Aggregated evaluations will be summarised in: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/OPTML-Group__NPO-WMDP__wmdp_bio/LMEval_SUMMARY.json
+[2026-05-29 21:58:28,713][lm_eval.evaluator][INFO] - Setting random seed to 0 | Setting numpy seed to 1234 | Setting torch manual seed to 1234 | Setting fewshot manual seed to 1234
+[2026-05-29 21:58:28,714][lm_eval.evaluator][INFO] - Using pre-initialized model
+[2026-05-29 21:58:35,564][lm_eval.tasks][INFO] - Selected tasks:
+[2026-05-29 21:58:35,564][lm_eval.tasks][INFO] - Task: wmdp_bio (wmdp/wmdp_bio.yaml)
+[2026-05-29 21:58:35,565][lm_eval.api.task][INFO] - Building contexts for wmdp_bio on rank 0...
+
  0%|          | 0/1273 [00:00<?, ?it/s]
  8%|▊         | 97/1273 [00:00<00:01, 966.69it/s]
 15%|█▌        | 194/1273 [00:00<00:01, 920.81it/s]
 23%|██▎       | 287/1273 [00:00<00:01, 906.76it/s]
 30%|██▉       | 378/1273 [00:00<00:00, 899.14it/s]
 37%|███▋      | 468/1273 [00:00<00:00, 892.36it/s]
 44%|████▍     | 558/1273 [00:00<00:00, 889.79it/s]
 51%|█████     | 647/1273 [00:00<00:00, 887.53it/s]
 58%|█████▊    | 736/1273 [00:00<00:00, 887.39it/s]
 65%|██████▍   | 825/1273 [00:00<00:00, 887.89it/s]
 72%|███████▏  | 915/1273 [00:01<00:00, 889.86it/s]
 79%|███████▉  | 1012/1273 [00:01<00:00, 912.66it/s]
 87%|████████▋ | 1111/1273 [00:01<00:00, 934.45it/s]
 95%|█████████▍| 1205/1273 [00:01<00:00, 929.65it/s]
+[2026-05-29 21:58:36,990][lm_eval.evaluator][INFO] - Running loglikelihood requests
+

data/ulars/supplement/standard_eval/logs/OPTML-Group__NPO-WMDP__wmdp_cyber.log

@@ -0,0 +1,22 @@
+repo_id: OPTML-Group/NPO-WMDP
+split: cyber
+snapshot: /zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/OPTML-Group__NPO-WMDP/models--OPTML-Group--NPO-WMDP/snapshots/fec7960fb3925037a6befc5819036cbc55664b18
+cmd: /zfspool/zangyihe/conda_envs/corgo/bin/python src/eval.py experiment=eval/wmdp/default model=zephyr-7b-beta model.model_args.pretrained_model_name_or_path=/zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/OPTML-Group__NPO-WMDP/models--OPTML-Group--NPO-WMDP/snapshots/fec7960fb3925037a6befc5819036cbc55664b18 model.tokenizer_args.pretrained_model_name_or_path=/zfspool/zangyihe/ulars/models/zephyr-7b-beta model.model_args.attn_implementation=sdpa model.model_args.torch_dtype=bfloat16 data_split=cyber task_name=hf_wmdp_cyber_OPTML-Group__NPO-WMDP paths.output_dir=/zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/OPTML-Group__NPO-WMDP__wmdp_cyber eval.lm_eval.overwrite=true eval.lm_eval.tasks=[wmdp_cyber] eval.lm_eval.simple_evaluate_args.batch_size=16
+
+/zfspool/zangyihe/conda_envs/corgo/lib/python3.11/site-packages/transformers/utils/hub.py:105: FutureWarning: Using `TRANSFORMERS_CACHE` is deprecated and will be removed in v5 of Transformers. Use `HF_HOME` instead.
+  warnings.warn(
+[2026-05-30 00:10:29,308] [INFO] [real_accelerator.py:219:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+
+[2026-05-30 00:10:45,089][lm_eval.models.huggingface][WARNING] - `pretrained` model kwarg is not of type `str`. Many other model arguments may be ignored. Please do not launch via accelerate or use `parallelize=True` if passing an existing model this way.
+[2026-05-30 00:10:45,101][lm_eval.models.huggingface][WARNING] - Passed an already-initialized model through `pretrained`, assuming single-process call to evaluate() or custom distributed integration
+[2026-05-30 00:10:45,102][evaluator][INFO] - ***** Running LMEval evaluation suite *****
+[2026-05-30 00:10:45,102][evaluator][INFO] - Fine-grained evaluations will be saved to: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/OPTML-Group__NPO-WMDP__wmdp_cyber/LMEval_EVAL.json
+[2026-05-30 00:10:45,102][evaluator][INFO] - Aggregated evaluations will be summarised in: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/OPTML-Group__NPO-WMDP__wmdp_cyber/LMEval_SUMMARY.json
+[2026-05-30 00:10:45,103][lm_eval.evaluator][INFO] - Setting random seed to 0 | Setting numpy seed to 1234 | Setting torch manual seed to 1234 | Setting fewshot manual seed to 1234
+[2026-05-30 00:10:45,103][lm_eval.evaluator][INFO] - Using pre-initialized model
+[2026-05-30 00:10:50,120][lm_eval.tasks][INFO] - Selected tasks:
+[2026-05-30 00:10:50,120][lm_eval.tasks][INFO] - Task: wmdp_cyber (wmdp/wmdp_cyber.yaml)
+[2026-05-30 00:10:50,120][lm_eval.api.task][INFO] - Building contexts for wmdp_cyber on rank 0...
+
  0%|          | 0/1987 [00:00<?, ?it/s]
  5%|▍         | 96/1987 [00:00<00:01, 955.60it/s]
 10%|▉         | 194/1987 [00:00<00:01, 964.49it/s]
 15%|█▍        | 292/1987 [00:00<00:01, 970.81it/s]
 20%|█▉        | 391/1987 [00:00<00:01, 974.55it/s]
 25%|██▍       | 490/1987 [00:00<00:01, 977.26it/s]
 30%|██▉       | 588/1987 [00:00<00:01, 976.39it/s]
 35%|███▍      | 687/1987 [00:00<00:01, 978.27it/s]
 40%|███▉      | 785/1987 [00:00<00:01, 974.08it/s]
 44%|████▍     | 884/1987 [00:00<00:01, 978.89it/s]
 50%|████▉     | 984/1987 [00:01<00:01, 983.08it/s]
 55%|█████▍    | 1083/1987 [00:01<00:00, 985.18it/s]
 60%|█████▉    | 1183/1987 [00:01<00:00, 986.77it/s]
 65%|██████▍   | 1283/1987 [00:01<00:00, 988.57it/s]
 70%|██████▉   | 1383/1987 [00:01<00:00, 990.20it/s]
 75%|███████▍  | 1483/1987 [00:01<00:00, 992.04it/s]
 80%|███████▉  | 1583/1987 [00:01<00:00, 992.02it/s]
 85%|████████▍ | 1683/1987 [00:01<00:00, 992.53it/s]
 90%|████████▉ | 1783/1987 [00:01<00:00, 987.44it/s]
 95%|█████████▍| 1882/1987 [00:01<00:00, 986.71it/s]
+[2026-05-30 00:10:52,191][lm_eval.evaluator][INFO] - Running loglikelihood requests
+

data/ulars/supplement/standard_eval/logs/aypan17__zephyr-7b-beta_cyber-unlearned__wmdp_bio.log

@@ -0,0 +1,22 @@
+repo_id: aypan17/zephyr-7b-beta_cyber-unlearned
+split: bio
+snapshot: /zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/aypan17__zephyr-7b-beta_cyber-unlearned/models--aypan17--zephyr-7b-beta_cyber-unlearned/snapshots/45a9456652ba1d90294c0d437dfb2d14542e2304
+cmd: /zfspool/zangyihe/conda_envs/corgo/bin/python src/eval.py experiment=eval/wmdp/default model=zephyr-7b-beta model.model_args.pretrained_model_name_or_path=/zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/aypan17__zephyr-7b-beta_cyber-unlearned/models--aypan17--zephyr-7b-beta_cyber-unlearned/snapshots/45a9456652ba1d90294c0d437dfb2d14542e2304 model.tokenizer_args.pretrained_model_name_or_path=/zfspool/zangyihe/ulars/models/zephyr-7b-beta model.model_args.attn_implementation=sdpa model.model_args.torch_dtype=bfloat16 data_split=bio task_name=hf_wmdp_bio_aypan17__zephyr-7b-beta_cyber-unlearned paths.output_dir=/zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/aypan17__zephyr-7b-beta_cyber-unlearned__wmdp_bio eval.lm_eval.overwrite=true eval.lm_eval.tasks=[wmdp_bio] eval.lm_eval.simple_evaluate_args.batch_size=16
+
+/zfspool/zangyihe/conda_envs/corgo/lib/python3.11/site-packages/transformers/utils/hub.py:105: FutureWarning: Using `TRANSFORMERS_CACHE` is deprecated and will be removed in v5 of Transformers. Use `HF_HOME` instead.
+  warnings.warn(
+[2026-05-29 21:58:19,978] [INFO] [real_accelerator.py:219:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+
+[2026-05-29 21:58:59,162][lm_eval.models.huggingface][WARNING] - `pretrained` model kwarg is not of type `str`. Many other model arguments may be ignored. Please do not launch via accelerate or use `parallelize=True` if passing an existing model this way.
+[2026-05-29 21:58:59,175][lm_eval.models.huggingface][WARNING] - Passed an already-initialized model through `pretrained`, assuming single-process call to evaluate() or custom distributed integration
+[2026-05-29 21:58:59,176][evaluator][INFO] - ***** Running LMEval evaluation suite *****
+[2026-05-29 21:58:59,176][evaluator][INFO] - Fine-grained evaluations will be saved to: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/aypan17__zephyr-7b-beta_cyber-unlearned__wmdp_bio/LMEval_EVAL.json
+[2026-05-29 21:58:59,176][evaluator][INFO] - Aggregated evaluations will be summarised in: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/aypan17__zephyr-7b-beta_cyber-unlearned__wmdp_bio/LMEval_SUMMARY.json
+[2026-05-29 21:58:59,178][lm_eval.evaluator][INFO] - Setting random seed to 0 | Setting numpy seed to 1234 | Setting torch manual seed to 1234 | Setting fewshot manual seed to 1234
+[2026-05-29 21:58:59,178][lm_eval.evaluator][INFO] - Using pre-initialized model
+[2026-05-29 21:59:07,970][lm_eval.tasks][INFO] - Selected tasks:
+[2026-05-29 21:59:07,970][lm_eval.tasks][INFO] - Task: wmdp_bio (wmdp/wmdp_bio.yaml)
+[2026-05-29 21:59:07,971][lm_eval.api.task][INFO] - Building contexts for wmdp_bio on rank 0...
+
  0%|          | 0/1273 [00:00<?, ?it/s]
  7%|▋         | 92/1273 [00:00<00:01, 918.51it/s]
 15%|█▍        | 188/1273 [00:00<00:01, 941.74it/s]
 22%|██▏       | 283/1273 [00:00<00:01, 873.65it/s]
 29%|██▉       | 373/1273 [00:00<00:01, 881.08it/s]
 37%|███▋      | 467/1273 [00:00<00:00, 899.10it/s]
 44%|████▍     | 562/1273 [00:00<00:00, 915.06it/s]
 52%|█████▏    | 657/1273 [00:00<00:00, 924.71it/s]
 59%|█████▉    | 754/1273 [00:00<00:00, 936.42it/s]
 67%|██████▋   | 850/1273 [00:00<00:00, 942.89it/s]
 74%|███████▍  | 945/1273 [00:01<00:00, 935.85it/s]
 82%|████████▏ | 1039/1273 [00:01<00:00, 932.76it/s]
 89%|████████▉ | 1137/1273 [00:01<00:00, 944.38it/s]
 97%|█████████▋| 1233/1273 [00:01<00:00, 948.08it/s]
+[2026-05-29 21:59:09,370][lm_eval.evaluator][INFO] - Running loglikelihood requests
+

data/ulars/supplement/standard_eval/logs/aypan17__zephyr-7b-beta_cyber-unlearned__wmdp_cyber.log

@@ -0,0 +1,22 @@
+repo_id: aypan17/zephyr-7b-beta_cyber-unlearned
+split: cyber
+snapshot: /zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/aypan17__zephyr-7b-beta_cyber-unlearned/models--aypan17--zephyr-7b-beta_cyber-unlearned/snapshots/45a9456652ba1d90294c0d437dfb2d14542e2304
+cmd: /zfspool/zangyihe/conda_envs/corgo/bin/python src/eval.py experiment=eval/wmdp/default model=zephyr-7b-beta model.model_args.pretrained_model_name_or_path=/zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/aypan17__zephyr-7b-beta_cyber-unlearned/models--aypan17--zephyr-7b-beta_cyber-unlearned/snapshots/45a9456652ba1d90294c0d437dfb2d14542e2304 model.tokenizer_args.pretrained_model_name_or_path=/zfspool/zangyihe/ulars/models/zephyr-7b-beta model.model_args.attn_implementation=sdpa model.model_args.torch_dtype=bfloat16 data_split=cyber task_name=hf_wmdp_cyber_aypan17__zephyr-7b-beta_cyber-unlearned paths.output_dir=/zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/aypan17__zephyr-7b-beta_cyber-unlearned__wmdp_cyber eval.lm_eval.overwrite=true eval.lm_eval.tasks=[wmdp_cyber] eval.lm_eval.simple_evaluate_args.batch_size=16
+
+/zfspool/zangyihe/conda_envs/corgo/lib/python3.11/site-packages/transformers/utils/hub.py:105: FutureWarning: Using `TRANSFORMERS_CACHE` is deprecated and will be removed in v5 of Transformers. Use `HF_HOME` instead.
+  warnings.warn(
+[2026-05-30 00:12:53,713] [INFO] [real_accelerator.py:219:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+
+[2026-05-30 00:13:01,380][lm_eval.models.huggingface][WARNING] - `pretrained` model kwarg is not of type `str`. Many other model arguments may be ignored. Please do not launch via accelerate or use `parallelize=True` if passing an existing model this way.
+[2026-05-30 00:13:01,393][lm_eval.models.huggingface][WARNING] - Passed an already-initialized model through `pretrained`, assuming single-process call to evaluate() or custom distributed integration
+[2026-05-30 00:13:01,394][evaluator][INFO] - ***** Running LMEval evaluation suite *****
+[2026-05-30 00:13:01,394][evaluator][INFO] - Fine-grained evaluations will be saved to: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/aypan17__zephyr-7b-beta_cyber-unlearned__wmdp_cyber/LMEval_EVAL.json
+[2026-05-30 00:13:01,394][evaluator][INFO] - Aggregated evaluations will be summarised in: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/aypan17__zephyr-7b-beta_cyber-unlearned__wmdp_cyber/LMEval_SUMMARY.json
+[2026-05-30 00:13:01,395][lm_eval.evaluator][INFO] - Setting random seed to 0 | Setting numpy seed to 1234 | Setting torch manual seed to 1234 | Setting fewshot manual seed to 1234
+[2026-05-30 00:13:01,396][lm_eval.evaluator][INFO] - Using pre-initialized model
+[2026-05-30 00:13:06,697][lm_eval.tasks][INFO] - Selected tasks:
+[2026-05-30 00:13:06,698][lm_eval.tasks][INFO] - Task: wmdp_cyber (wmdp/wmdp_cyber.yaml)
+[2026-05-30 00:13:06,698][lm_eval.api.task][INFO] - Building contexts for wmdp_cyber on rank 0...
+
  0%|          | 0/1987 [00:00<?, ?it/s]
  5%|▍         | 96/1987 [00:00<00:01, 955.44it/s]
 10%|▉         | 194/1987 [00:00<00:01, 968.10it/s]
 15%|█▍        | 293/1987 [00:00<00:01, 975.89it/s]
 20%|█▉        | 392/1987 [00:00<00:01, 979.56it/s]
 25%|██▍       | 491/1987 [00:00<00:01, 982.16it/s]
 30%|██▉       | 591/1987 [00:00<00:01, 985.75it/s]
 35%|███▍      | 690/1987 [00:00<00:01, 979.64it/s]
 40%|███▉      | 790/1987 [00:00<00:01, 983.96it/s]
 45%|████▍     | 890/1987 [00:00<00:01, 985.97it/s]
 50%|████▉     | 989/1987 [00:01<00:01, 986.01it/s]
 55%|█████▍    | 1089/1987 [00:01<00:00, 987.83it/s]
 60%|█████▉    | 1188/1987 [00:01<00:00, 978.58it/s]
 65%|██████▍   | 1287/1987 [00:01<00:00, 981.43it/s]
 70%|██████▉   | 1387/1987 [00:01<00:00, 984.43it/s]
 75%|███████▍  | 1486/1987 [00:01<00:00, 980.83it/s]
 80%|███████▉  | 1586/1987 [00:01<00:00, 985.06it/s]
 85%|████████▍ | 1686/1987 [00:01<00:00, 986.61it/s]
 90%|████████▉ | 1786/1987 [00:01<00:00, 988.07it/s]
 95%|█████████▍| 1885/1987 [00:01<00:00, 987.92it/s]
+[2026-05-30 00:13:08,768][lm_eval.evaluator][INFO] - Running loglikelihood requests
+

data/ulars/supplement/standard_eval/logs/concept-unlearning__zephyr-7b-beta_ft_lora_all_novels_v3_ft_npo_gdr_lora_positive_dataset_v3__wmdp_bio.log

@@ -0,0 +1,22 @@
+repo_id: concept-unlearning/zephyr-7b-beta_ft_lora_all_novels_v3_ft_npo_gdr_lora_positive_dataset_v3
+split: bio
+snapshot: /zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/concept-unlearning__zephyr-7b-beta_ft_lora_all_novels_v3_ft_npo_gdr_lora_positive_dataset_v3/models--concept-unlearning--zephyr-7b-beta_ft_lora_all_novels_v3_ft_npo_gdr_lora_positive_dataset_v3/snapshots/39ea9fc05165adc743467d89dfffbd00576ff8c1
+cmd: /zfspool/zangyihe/conda_envs/corgo/bin/python src/eval.py experiment=eval/wmdp/default model=zephyr-7b-beta model.model_args.pretrained_model_name_or_path=/zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/concept-unlearning__zephyr-7b-beta_ft_lora_all_novels_v3_ft_npo_gdr_lora_positive_dataset_v3/models--concept-unlearning--zephyr-7b-beta_ft_lora_all_novels_v3_ft_npo_gdr_lora_positive_dataset_v3/snapshots/39ea9fc05165adc743467d89dfffbd00576ff8c1 model.tokenizer_args.pretrained_model_name_or_path=/zfspool/zangyihe/ulars/models/zephyr-7b-beta model.model_args.attn_implementation=sdpa model.model_args.torch_dtype=bfloat16 data_split=bio task_name=hf_wmdp_bio_concept-unlearning__zephyr-7b-beta_ft_lora_all_novels_v3_ft_npo_gdr_lora_positive_dataset_v3 paths.output_dir=/zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/concept-unlearning__zephyr-7b-beta_ft_lora_all_novels_v3_ft_npo_gdr_lora_positive_dataset_v3__wmdp_bio eval.lm_eval.overwrite=true eval.lm_eval.tasks=[wmdp_bio] eval.lm_eval.simple_evaluate_args.batch_size=16
+
+/zfspool/zangyihe/conda_envs/corgo/lib/python3.11/site-packages/transformers/utils/hub.py:105: FutureWarning: Using `TRANSFORMERS_CACHE` is deprecated and will be removed in v5 of Transformers. Use `HF_HOME` instead.
+  warnings.warn(
+[2026-05-29 23:52:11,204] [INFO] [real_accelerator.py:219:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+
+[2026-05-29 23:52:46,787][lm_eval.models.huggingface][WARNING] - `pretrained` model kwarg is not of type `str`. Many other model arguments may be ignored. Please do not launch via accelerate or use `parallelize=True` if passing an existing model this way.
+[2026-05-29 23:52:46,801][lm_eval.models.huggingface][WARNING] - Passed an already-initialized model through `pretrained`, assuming single-process call to evaluate() or custom distributed integration
+[2026-05-29 23:52:46,802][evaluator][INFO] - ***** Running LMEval evaluation suite *****
+[2026-05-29 23:52:46,803][evaluator][INFO] - Fine-grained evaluations will be saved to: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/concept-unlearning__zephyr-7b-beta_ft_lora_all_novels_v3_ft_npo_gdr_lora_positive_dataset_v3__wmdp_bio/LMEval_EVAL.json
+[2026-05-29 23:52:46,803][evaluator][INFO] - Aggregated evaluations will be summarised in: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/concept-unlearning__zephyr-7b-beta_ft_lora_all_novels_v3_ft_npo_gdr_lora_positive_dataset_v3__wmdp_bio/LMEval_SUMMARY.json
+[2026-05-29 23:52:46,804][lm_eval.evaluator][INFO] - Setting random seed to 0 | Setting numpy seed to 1234 | Setting torch manual seed to 1234 | Setting fewshot manual seed to 1234
+[2026-05-29 23:52:46,804][lm_eval.evaluator][INFO] - Using pre-initialized model
+[2026-05-29 23:52:52,232][lm_eval.tasks][INFO] - Selected tasks:
+[2026-05-29 23:52:52,232][lm_eval.tasks][INFO] - Task: wmdp_bio (wmdp/wmdp_bio.yaml)
+[2026-05-29 23:52:52,232][lm_eval.api.task][INFO] - Building contexts for wmdp_bio on rank 0...
+
  0%|          | 0/1273 [00:00<?, ?it/s]
  7%|▋         | 95/1273 [00:00<00:01, 940.27it/s]
 15%|█▌        | 192/1273 [00:00<00:01, 954.54it/s]
 23%|██▎       | 290/1273 [00:00<00:01, 961.93it/s]
 31%|███       | 389/1273 [00:00<00:00, 969.12it/s]
 38%|███▊      | 488/1273 [00:00<00:00, 973.50it/s]
 46%|████▌     | 587/1273 [00:00<00:00, 977.79it/s]
 54%|█████▍    | 685/1273 [00:00<00:00, 968.92it/s]
 62%|██████▏   | 784/1273 [00:00<00:00, 972.65it/s]
 69%|██████▉   | 883/1273 [00:00<00:00, 975.26it/s]
 77%|███████▋  | 982/1273 [00:01<00:00, 977.15it/s]
 85%|████████▍ | 1081/1273 [00:01<00:00, 979.59it/s]
 93%|█████████▎| 1179/1273 [00:01<00:00, 965.72it/s]
+[2026-05-29 23:52:53,576][lm_eval.evaluator][INFO] - Running loglikelihood requests
+

data/ulars/supplement/standard_eval/logs/concept-unlearning__zephyr-7b-beta_ft_lora_all_novels_v3_ft_npo_gdr_lora_positive_dataset_v3__wmdp_cyber.log

@@ -0,0 +1,22 @@
+repo_id: concept-unlearning/zephyr-7b-beta_ft_lora_all_novels_v3_ft_npo_gdr_lora_positive_dataset_v3
+split: cyber
+snapshot: /zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/concept-unlearning__zephyr-7b-beta_ft_lora_all_novels_v3_ft_npo_gdr_lora_positive_dataset_v3/models--concept-unlearning--zephyr-7b-beta_ft_lora_all_novels_v3_ft_npo_gdr_lora_positive_dataset_v3/snapshots/39ea9fc05165adc743467d89dfffbd00576ff8c1
+cmd: /zfspool/zangyihe/conda_envs/corgo/bin/python src/eval.py experiment=eval/wmdp/default model=zephyr-7b-beta model.model_args.pretrained_model_name_or_path=/zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/concept-unlearning__zephyr-7b-beta_ft_lora_all_novels_v3_ft_npo_gdr_lora_positive_dataset_v3/models--concept-unlearning--zephyr-7b-beta_ft_lora_all_novels_v3_ft_npo_gdr_lora_positive_dataset_v3/snapshots/39ea9fc05165adc743467d89dfffbd00576ff8c1 model.tokenizer_args.pretrained_model_name_or_path=/zfspool/zangyihe/ulars/models/zephyr-7b-beta model.model_args.attn_implementation=sdpa model.model_args.torch_dtype=bfloat16 data_split=cyber task_name=hf_wmdp_cyber_concept-unlearning__zephyr-7b-beta_ft_lora_all_novels_v3_ft_npo_gdr_lora_positive_dataset_v3 paths.output_dir=/zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/concept-unlearning__zephyr-7b-beta_ft_lora_all_novels_v3_ft_npo_gdr_lora_positive_dataset_v3__wmdp_cyber eval.lm_eval.overwrite=true eval.lm_eval.tasks=[wmdp_cyber] eval.lm_eval.simple_evaluate_args.batch_size=16
+
+/zfspool/zangyihe/conda_envs/corgo/lib/python3.11/site-packages/transformers/utils/hub.py:105: FutureWarning: Using `TRANSFORMERS_CACHE` is deprecated and will be removed in v5 of Transformers. Use `HF_HOME` instead.
+  warnings.warn(
+[2026-05-30 00:38:12,289] [INFO] [real_accelerator.py:219:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+
+[2026-05-30 00:38:22,247][lm_eval.models.huggingface][WARNING] - `pretrained` model kwarg is not of type `str`. Many other model arguments may be ignored. Please do not launch via accelerate or use `parallelize=True` if passing an existing model this way.
+[2026-05-30 00:38:22,260][lm_eval.models.huggingface][WARNING] - Passed an already-initialized model through `pretrained`, assuming single-process call to evaluate() or custom distributed integration
+[2026-05-30 00:38:22,262][evaluator][INFO] - ***** Running LMEval evaluation suite *****
+[2026-05-30 00:38:22,262][evaluator][INFO] - Fine-grained evaluations will be saved to: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/concept-unlearning__zephyr-7b-beta_ft_lora_all_novels_v3_ft_npo_gdr_lora_positive_dataset_v3__wmdp_cyber/LMEval_EVAL.json
+[2026-05-30 00:38:22,262][evaluator][INFO] - Aggregated evaluations will be summarised in: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/concept-unlearning__zephyr-7b-beta_ft_lora_all_novels_v3_ft_npo_gdr_lora_positive_dataset_v3__wmdp_cyber/LMEval_SUMMARY.json
+[2026-05-30 00:38:22,264][lm_eval.evaluator][INFO] - Setting random seed to 0 | Setting numpy seed to 1234 | Setting torch manual seed to 1234 | Setting fewshot manual seed to 1234
+[2026-05-30 00:38:22,264][lm_eval.evaluator][INFO] - Using pre-initialized model
+[2026-05-30 00:38:27,316][lm_eval.tasks][INFO] - Selected tasks:
+[2026-05-30 00:38:27,316][lm_eval.tasks][INFO] - Task: wmdp_cyber (wmdp/wmdp_cyber.yaml)
+[2026-05-30 00:38:27,317][lm_eval.api.task][INFO] - Building contexts for wmdp_cyber on rank 0...
+
  0%|          | 0/1987 [00:00<?, ?it/s]
  5%|▍         | 94/1987 [00:00<00:02, 934.44it/s]
 10%|▉         | 192/1987 [00:00<00:01, 959.90it/s]
 15%|█▍        | 289/1987 [00:00<00:01, 960.73it/s]
 20%|█▉        | 388/1987 [00:00<00:01, 970.80it/s]
 25%|██▍       | 488/1987 [00:00<00:01, 979.67it/s]
 29%|██▉       | 586/1987 [00:00<00:01, 973.96it/s]
 35%|███▍      | 687/1987 [00:00<00:01, 982.66it/s]
 40%|███▉      | 788/1987 [00:00<00:01, 988.46it/s]
 45%|████▍     | 887/1987 [00:00<00:01, 982.80it/s]
 50%|████▉     | 987/1987 [00:01<00:01, 987.92it/s]
 55%|█████▍    | 1087/1987 [00:01<00:00, 991.15it/s]
 60%|█████▉    | 1187/1987 [00:01<00:00, 979.69it/s]
 65%|██████▍   | 1287/1987 [00:01<00:00, 984.92it/s]
 70%|██████▉   | 1388/1987 [00:01<00:00, 989.74it/s]
 75%|███████▍  | 1489/1987 [00:01<00:00, 992.95it/s]
 80%|███████▉  | 1589/1987 [00:01<00:00, 981.19it/s]
 85%|████████▌ | 1690/1987 [00:01<00:00, 986.92it/s]
 90%|█████████ | 1791/1987 [00:01<00:00, 990.93it/s]
 95%|█████████▌| 1892/1987 [00:01<00:00, 994.21it/s]
+[2026-05-30 00:38:29,384][lm_eval.evaluator][INFO] - Running loglikelihood requests
+

data/ulars/supplement/standard_eval/logs/concept-unlearning__zephyr-7b-beta_ft_lora_all_novels_v3_ft_rmu_lora_positive_dataset_v1__wmdp_bio.log

@@ -0,0 +1,22 @@
+repo_id: concept-unlearning/zephyr-7b-beta_ft_lora_all_novels_v3_ft_rmu_lora_positive_dataset_v1
+split: bio
+snapshot: /zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/concept-unlearning__zephyr-7b-beta_ft_lora_all_novels_v3_ft_rmu_lora_positive_dataset_v1/models--concept-unlearning--zephyr-7b-beta_ft_lora_all_novels_v3_ft_rmu_lora_positive_dataset_v1/snapshots/546bc212b9f71ab1137534d4125741911afddcdc
+cmd: /zfspool/zangyihe/conda_envs/corgo/bin/python src/eval.py experiment=eval/wmdp/default model=zephyr-7b-beta model.model_args.pretrained_model_name_or_path=/zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/concept-unlearning__zephyr-7b-beta_ft_lora_all_novels_v3_ft_rmu_lora_positive_dataset_v1/models--concept-unlearning--zephyr-7b-beta_ft_lora_all_novels_v3_ft_rmu_lora_positive_dataset_v1/snapshots/546bc212b9f71ab1137534d4125741911afddcdc model.tokenizer_args.pretrained_model_name_or_path=/zfspool/zangyihe/ulars/models/zephyr-7b-beta model.model_args.attn_implementation=sdpa model.model_args.torch_dtype=bfloat16 data_split=bio task_name=hf_wmdp_bio_concept-unlearning__zephyr-7b-beta_ft_lora_all_novels_v3_ft_rmu_lora_positive_dataset_v1 paths.output_dir=/zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/concept-unlearning__zephyr-7b-beta_ft_lora_all_novels_v3_ft_rmu_lora_positive_dataset_v1__wmdp_bio eval.lm_eval.overwrite=true eval.lm_eval.tasks=[wmdp_bio] eval.lm_eval.simple_evaluate_args.batch_size=16
+
+/zfspool/zangyihe/conda_envs/corgo/lib/python3.11/site-packages/transformers/utils/hub.py:105: FutureWarning: Using `TRANSFORMERS_CACHE` is deprecated and will be removed in v5 of Transformers. Use `HF_HOME` instead.
+  warnings.warn(
+[2026-05-29 22:51:31,387] [INFO] [real_accelerator.py:219:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+
+[2026-05-29 22:51:55,487][lm_eval.models.huggingface][WARNING] - `pretrained` model kwarg is not of type `str`. Many other model arguments may be ignored. Please do not launch via accelerate or use `parallelize=True` if passing an existing model this way.
+[2026-05-29 22:51:55,499][lm_eval.models.huggingface][WARNING] - Passed an already-initialized model through `pretrained`, assuming single-process call to evaluate() or custom distributed integration
+[2026-05-29 22:51:55,500][evaluator][INFO] - ***** Running LMEval evaluation suite *****
+[2026-05-29 22:51:55,500][evaluator][INFO] - Fine-grained evaluations will be saved to: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/concept-unlearning__zephyr-7b-beta_ft_lora_all_novels_v3_ft_rmu_lora_positive_dataset_v1__wmdp_bio/LMEval_EVAL.json
+[2026-05-29 22:51:55,500][evaluator][INFO] - Aggregated evaluations will be summarised in: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/concept-unlearning__zephyr-7b-beta_ft_lora_all_novels_v3_ft_rmu_lora_positive_dataset_v1__wmdp_bio/LMEval_SUMMARY.json
+[2026-05-29 22:51:55,501][lm_eval.evaluator][INFO] - Setting random seed to 0 | Setting numpy seed to 1234 | Setting torch manual seed to 1234 | Setting fewshot manual seed to 1234
+[2026-05-29 22:51:55,501][lm_eval.evaluator][INFO] - Using pre-initialized model
+[2026-05-29 22:52:09,174][lm_eval.tasks][INFO] - Selected tasks:
+[2026-05-29 22:52:09,175][lm_eval.tasks][INFO] - Task: wmdp_bio (wmdp/wmdp_bio.yaml)
+[2026-05-29 22:52:09,175][lm_eval.api.task][INFO] - Building contexts for wmdp_bio on rank 0...
+
  0%|          | 0/1273 [00:00<?, ?it/s]
  8%|▊         | 96/1273 [00:00<00:01, 959.78it/s]
 15%|█▌        | 194/1273 [00:00<00:01, 970.65it/s]
 23%|██▎       | 293/1273 [00:00<00:01, 976.80it/s]
 31%|███       | 392/1273 [00:00<00:00, 981.19it/s]
 39%|███▊      | 491/1273 [00:00<00:00, 980.62it/s]
 46%|████▋     | 590/1273 [00:00<00:00, 982.87it/s]
 54%|█████▍    | 689/1273 [00:00<00:00, 983.47it/s]
 62%|██████▏   | 788/1273 [00:00<00:00, 985.50it/s]
 70%|██████▉   | 888/1273 [00:00<00:00, 989.14it/s]
 78%|███████▊  | 987/1273 [00:01<00:00, 979.10it/s]
 85%|████████▌ | 1086/1273 [00:01<00:00, 980.96it/s]
 93%|█████████▎| 1185/1273 [00:01<00:00, 981.55it/s]
+[2026-05-29 22:52:10,503][lm_eval.evaluator][INFO] - Running loglikelihood requests
+

data/ulars/supplement/standard_eval/logs/concept-unlearning__zephyr-7b-beta_ft_lora_all_novels_v3_ft_rmu_lora_positive_dataset_v1__wmdp_cyber.log

@@ -0,0 +1,22 @@
+repo_id: concept-unlearning/zephyr-7b-beta_ft_lora_all_novels_v3_ft_rmu_lora_positive_dataset_v1
+split: cyber
+snapshot: /zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/concept-unlearning__zephyr-7b-beta_ft_lora_all_novels_v3_ft_rmu_lora_positive_dataset_v1/models--concept-unlearning--zephyr-7b-beta_ft_lora_all_novels_v3_ft_rmu_lora_positive_dataset_v1/snapshots/546bc212b9f71ab1137534d4125741911afddcdc
+cmd: /zfspool/zangyihe/conda_envs/corgo/bin/python src/eval.py experiment=eval/wmdp/default model=zephyr-7b-beta model.model_args.pretrained_model_name_or_path=/zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/concept-unlearning__zephyr-7b-beta_ft_lora_all_novels_v3_ft_rmu_lora_positive_dataset_v1/models--concept-unlearning--zephyr-7b-beta_ft_lora_all_novels_v3_ft_rmu_lora_positive_dataset_v1/snapshots/546bc212b9f71ab1137534d4125741911afddcdc model.tokenizer_args.pretrained_model_name_or_path=/zfspool/zangyihe/ulars/models/zephyr-7b-beta model.model_args.attn_implementation=sdpa model.model_args.torch_dtype=bfloat16 data_split=cyber task_name=hf_wmdp_cyber_concept-unlearning__zephyr-7b-beta_ft_lora_all_novels_v3_ft_rmu_lora_positive_dataset_v1 paths.output_dir=/zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/concept-unlearning__zephyr-7b-beta_ft_lora_all_novels_v3_ft_rmu_lora_positive_dataset_v1__wmdp_cyber eval.lm_eval.overwrite=true eval.lm_eval.tasks=[wmdp_cyber] eval.lm_eval.simple_evaluate_args.batch_size=16
+
+/zfspool/zangyihe/conda_envs/corgo/lib/python3.11/site-packages/transformers/utils/hub.py:105: FutureWarning: Using `TRANSFORMERS_CACHE` is deprecated and will be removed in v5 of Transformers. Use `HF_HOME` instead.
+  warnings.warn(
+[2026-05-30 00:38:50,633] [INFO] [real_accelerator.py:219:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+
+[2026-05-30 00:38:58,732][lm_eval.models.huggingface][WARNING] - `pretrained` model kwarg is not of type `str`. Many other model arguments may be ignored. Please do not launch via accelerate or use `parallelize=True` if passing an existing model this way.
+[2026-05-30 00:38:58,745][lm_eval.models.huggingface][WARNING] - Passed an already-initialized model through `pretrained`, assuming single-process call to evaluate() or custom distributed integration
+[2026-05-30 00:38:58,746][evaluator][INFO] - ***** Running LMEval evaluation suite *****
+[2026-05-30 00:38:58,746][evaluator][INFO] - Fine-grained evaluations will be saved to: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/concept-unlearning__zephyr-7b-beta_ft_lora_all_novels_v3_ft_rmu_lora_positive_dataset_v1__wmdp_cyber/LMEval_EVAL.json
+[2026-05-30 00:38:58,746][evaluator][INFO] - Aggregated evaluations will be summarised in: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/concept-unlearning__zephyr-7b-beta_ft_lora_all_novels_v3_ft_rmu_lora_positive_dataset_v1__wmdp_cyber/LMEval_SUMMARY.json
+[2026-05-30 00:38:58,747][lm_eval.evaluator][INFO] - Setting random seed to 0 | Setting numpy seed to 1234 | Setting torch manual seed to 1234 | Setting fewshot manual seed to 1234
+[2026-05-30 00:38:58,747][lm_eval.evaluator][INFO] - Using pre-initialized model
+[2026-05-30 00:39:03,987][lm_eval.tasks][INFO] - Selected tasks:
+[2026-05-30 00:39:03,987][lm_eval.tasks][INFO] - Task: wmdp_cyber (wmdp/wmdp_cyber.yaml)
+[2026-05-30 00:39:03,987][lm_eval.api.task][INFO] - Building contexts for wmdp_cyber on rank 0...
+
  0%|          | 0/1987 [00:00<?, ?it/s]
  5%|▍         | 95/1987 [00:00<00:02, 944.56it/s]
 10%|▉         | 193/1987 [00:00<00:01, 960.10it/s]
 15%|█▍        | 290/1987 [00:00<00:01, 948.13it/s]
 20%|█▉        | 388/1987 [00:00<00:01, 959.24it/s]
 25%|██▍       | 487/1987 [00:00<00:01, 968.19it/s]
 29%|██▉       | 584/1987 [00:00<00:01, 967.11it/s]
 34%|███▍      | 683/1987 [00:00<00:01, 973.35it/s]
 39%|███▉      | 783/1987 [00:00<00:01, 979.25it/s]
 44%|████▍     | 882/1987 [00:00<00:01, 982.27it/s]
 49%|████▉     | 981/1987 [00:01<00:01, 984.60it/s]
 54%|█████▍    | 1080/1987 [00:01<00:00, 979.81it/s]
 59%|█████▉    | 1179/1987 [00:01<00:00, 981.34it/s]
 64%|██████▍   | 1278/1987 [00:01<00:00, 983.41it/s]
 69%|██████▉   | 1378/1987 [00:01<00:00, 985.87it/s]
 74%|███████▍  | 1478/1987 [00:01<00:00, 987.40it/s]
 79%|███████▉  | 1577/1987 [00:01<00:00, 987.12it/s]
 84%|████████▍ | 1677/1987 [00:01<00:00, 988.23it/s]
 89%|████████▉ | 1777/1987 [00:01<00:00, 989.88it/s]
 94%|█████████▍| 1877/1987 [00:01<00:00, 990.65it/s]
 99%|█████████▉| 1977/1987 [00:02<00:00, 991.19it/s]
+[2026-05-30 00:39:06,064][lm_eval.evaluator][INFO] - Running loglikelihood requests
+

data/ulars/supplement/standard_eval/logs/concept-unlearning__zephyr-7b-beta_ft_lora_civil_comments_v1_ft_npo_gdr_lora_toxic_v2__wmdp_bio.log

@@ -0,0 +1,22 @@
+repo_id: concept-unlearning/zephyr-7b-beta_ft_lora_civil_comments_v1_ft_npo_gdr_lora_toxic_v2
+split: bio
+snapshot: /zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/concept-unlearning__zephyr-7b-beta_ft_lora_civil_comments_v1_ft_npo_gdr_lora_toxic_v2/models--concept-unlearning--zephyr-7b-beta_ft_lora_civil_comments_v1_ft_npo_gdr_lora_toxic_v2/snapshots/49845d64094262c6b27c3508a1dc9fca41cb6998
+cmd: /zfspool/zangyihe/conda_envs/corgo/bin/python src/eval.py experiment=eval/wmdp/default model=zephyr-7b-beta model.model_args.pretrained_model_name_or_path=/zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/concept-unlearning__zephyr-7b-beta_ft_lora_civil_comments_v1_ft_npo_gdr_lora_toxic_v2/models--concept-unlearning--zephyr-7b-beta_ft_lora_civil_comments_v1_ft_npo_gdr_lora_toxic_v2/snapshots/49845d64094262c6b27c3508a1dc9fca41cb6998 model.tokenizer_args.pretrained_model_name_or_path=/zfspool/zangyihe/ulars/models/zephyr-7b-beta model.model_args.attn_implementation=sdpa model.model_args.torch_dtype=bfloat16 data_split=bio task_name=hf_wmdp_bio_concept-unlearning__zephyr-7b-beta_ft_lora_civil_comments_v1_ft_npo_gdr_lora_toxic_v2 paths.output_dir=/zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/concept-unlearning__zephyr-7b-beta_ft_lora_civil_comments_v1_ft_npo_gdr_lora_toxic_v2__wmdp_bio eval.lm_eval.overwrite=true eval.lm_eval.tasks=[wmdp_bio] eval.lm_eval.simple_evaluate_args.batch_size=16
+
+/zfspool/zangyihe/conda_envs/corgo/lib/python3.11/site-packages/transformers/utils/hub.py:105: FutureWarning: Using `TRANSFORMERS_CACHE` is deprecated and will be removed in v5 of Transformers. Use `HF_HOME` instead.
+  warnings.warn(
+[2026-05-29 22:58:20,721] [INFO] [real_accelerator.py:219:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+
+[2026-05-29 22:58:43,912][lm_eval.models.huggingface][WARNING] - `pretrained` model kwarg is not of type `str`. Many other model arguments may be ignored. Please do not launch via accelerate or use `parallelize=True` if passing an existing model this way.
+[2026-05-29 22:58:43,924][lm_eval.models.huggingface][WARNING] - Passed an already-initialized model through `pretrained`, assuming single-process call to evaluate() or custom distributed integration
+[2026-05-29 22:58:43,925][evaluator][INFO] - ***** Running LMEval evaluation suite *****
+[2026-05-29 22:58:43,925][evaluator][INFO] - Fine-grained evaluations will be saved to: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/concept-unlearning__zephyr-7b-beta_ft_lora_civil_comments_v1_ft_npo_gdr_lora_toxic_v2__wmdp_bio/LMEval_EVAL.json
+[2026-05-29 22:58:43,925][evaluator][INFO] - Aggregated evaluations will be summarised in: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/concept-unlearning__zephyr-7b-beta_ft_lora_civil_comments_v1_ft_npo_gdr_lora_toxic_v2__wmdp_bio/LMEval_SUMMARY.json
+[2026-05-29 22:58:43,926][lm_eval.evaluator][INFO] - Setting random seed to 0 | Setting numpy seed to 1234 | Setting torch manual seed to 1234 | Setting fewshot manual seed to 1234
+[2026-05-29 22:58:43,927][lm_eval.evaluator][INFO] - Using pre-initialized model
+[2026-05-29 22:58:52,876][lm_eval.tasks][INFO] - Selected tasks:
+[2026-05-29 22:58:52,876][lm_eval.tasks][INFO] - Task: wmdp_bio (wmdp/wmdp_bio.yaml)
+[2026-05-29 22:58:52,876][lm_eval.api.task][INFO] - Building contexts for wmdp_bio on rank 0...
+
  0%|          | 0/1273 [00:00<?, ?it/s]
  7%|▋         | 95/1273 [00:00<00:01, 940.52it/s]
 15%|█▌        | 194/1273 [00:00<00:01, 965.12it/s]
 23%|██▎       | 293/1273 [00:00<00:01, 974.04it/s]
 31%|███       | 393/1273 [00:00<00:00, 981.07it/s]
 39%|███▊      | 492/1273 [00:00<00:00, 984.12it/s]
 46%|████▋     | 591/1273 [00:00<00:00, 985.80it/s]
 54%|█████▍    | 691/1273 [00:00<00:00, 989.39it/s]
 62%|██████▏   | 791/1273 [00:00<00:00, 991.74it/s]
 70%|██████▉   | 891/1273 [00:00<00:00, 992.41it/s]
 78%|███████▊  | 991/1273 [00:01<00:00, 994.20it/s]
 86%|████████▌ | 1091/1273 [00:01<00:00, 992.55it/s]
 94%|█████████▎| 1191/1273 [00:01<00:00, 992.05it/s]
+[2026-05-29 22:58:54,196][lm_eval.evaluator][INFO] - Running loglikelihood requests
+

data/ulars/supplement/standard_eval/logs/concept-unlearning__zephyr-7b-beta_ft_lora_civil_comments_v1_ft_npo_gdr_lora_toxic_v2__wmdp_cyber.log

@@ -0,0 +1,22 @@
+repo_id: concept-unlearning/zephyr-7b-beta_ft_lora_civil_comments_v1_ft_npo_gdr_lora_toxic_v2
+split: cyber
+snapshot: /zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/concept-unlearning__zephyr-7b-beta_ft_lora_civil_comments_v1_ft_npo_gdr_lora_toxic_v2/models--concept-unlearning--zephyr-7b-beta_ft_lora_civil_comments_v1_ft_npo_gdr_lora_toxic_v2/snapshots/49845d64094262c6b27c3508a1dc9fca41cb6998
+cmd: /zfspool/zangyihe/conda_envs/corgo/bin/python src/eval.py experiment=eval/wmdp/default model=zephyr-7b-beta model.model_args.pretrained_model_name_or_path=/zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/concept-unlearning__zephyr-7b-beta_ft_lora_civil_comments_v1_ft_npo_gdr_lora_toxic_v2/models--concept-unlearning--zephyr-7b-beta_ft_lora_civil_comments_v1_ft_npo_gdr_lora_toxic_v2/snapshots/49845d64094262c6b27c3508a1dc9fca41cb6998 model.tokenizer_args.pretrained_model_name_or_path=/zfspool/zangyihe/ulars/models/zephyr-7b-beta model.model_args.attn_implementation=sdpa model.model_args.torch_dtype=bfloat16 data_split=cyber task_name=hf_wmdp_cyber_concept-unlearning__zephyr-7b-beta_ft_lora_civil_comments_v1_ft_npo_gdr_lora_toxic_v2 paths.output_dir=/zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/concept-unlearning__zephyr-7b-beta_ft_lora_civil_comments_v1_ft_npo_gdr_lora_toxic_v2__wmdp_cyber eval.lm_eval.overwrite=true eval.lm_eval.tasks=[wmdp_cyber] eval.lm_eval.simple_evaluate_args.batch_size=16
+
+/zfspool/zangyihe/conda_envs/corgo/lib/python3.11/site-packages/transformers/utils/hub.py:105: FutureWarning: Using `TRANSFORMERS_CACHE` is deprecated and will be removed in v5 of Transformers. Use `HF_HOME` instead.
+  warnings.warn(
+[2026-05-30 00:40:31,958] [INFO] [real_accelerator.py:219:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+
+[2026-05-30 00:40:39,900][lm_eval.models.huggingface][WARNING] - `pretrained` model kwarg is not of type `str`. Many other model arguments may be ignored. Please do not launch via accelerate or use `parallelize=True` if passing an existing model this way.
+[2026-05-30 00:40:39,913][lm_eval.models.huggingface][WARNING] - Passed an already-initialized model through `pretrained`, assuming single-process call to evaluate() or custom distributed integration
+[2026-05-30 00:40:39,914][evaluator][INFO] - ***** Running LMEval evaluation suite *****
+[2026-05-30 00:40:39,915][evaluator][INFO] - Fine-grained evaluations will be saved to: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/concept-unlearning__zephyr-7b-beta_ft_lora_civil_comments_v1_ft_npo_gdr_lora_toxic_v2__wmdp_cyber/LMEval_EVAL.json
+[2026-05-30 00:40:39,915][evaluator][INFO] - Aggregated evaluations will be summarised in: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/concept-unlearning__zephyr-7b-beta_ft_lora_civil_comments_v1_ft_npo_gdr_lora_toxic_v2__wmdp_cyber/LMEval_SUMMARY.json
+[2026-05-30 00:40:39,916][lm_eval.evaluator][INFO] - Setting random seed to 0 | Setting numpy seed to 1234 | Setting torch manual seed to 1234 | Setting fewshot manual seed to 1234
+[2026-05-30 00:40:39,916][lm_eval.evaluator][INFO] - Using pre-initialized model
+[2026-05-30 00:40:48,791][lm_eval.tasks][INFO] - Selected tasks:
+[2026-05-30 00:40:48,791][lm_eval.tasks][INFO] - Task: wmdp_cyber (wmdp/wmdp_cyber.yaml)
+[2026-05-30 00:40:48,791][lm_eval.api.task][INFO] - Building contexts for wmdp_cyber on rank 0...
+
  0%|          | 0/1987 [00:00<?, ?it/s]
  5%|▍         | 96/1987 [00:00<00:01, 951.08it/s]
 10%|▉         | 194/1987 [00:00<00:01, 967.40it/s]
 15%|█▍        | 293/1987 [00:00<00:01, 975.82it/s]
 20%|█▉        | 391/1987 [00:00<00:01, 974.75it/s]
 25%|██▍       | 490/1987 [00:00<00:01, 979.78it/s]
 30%|██▉       | 590/1987 [00:00<00:01, 983.70it/s]
 35%|███▍      | 689/1987 [00:00<00:01, 984.70it/s]
 40%|███▉      | 789/1987 [00:00<00:01, 987.92it/s]
 45%|████▍     | 888/1987 [00:00<00:01, 987.19it/s]
 50%|████▉     | 988/1987 [00:01<00:01, 989.11it/s]
 55%|█████▍    | 1088/1987 [00:01<00:00, 990.79it/s]
 60%|█████▉    | 1188/1987 [00:01<00:00, 989.99it/s]
 65%|██████▍   | 1287/1987 [00:01<00:00, 989.64it/s]
 70%|██████▉   | 1387/1987 [00:01<00:00, 991.34it/s]
 75%|███████▍  | 1487/1987 [00:01<00:00, 993.50it/s]
 80%|███████▉  | 1588/1987 [00:01<00:00, 995.95it/s]
 85%|████████▍ | 1688/1987 [00:01<00:00, 986.31it/s]
 90%|████████▉ | 1788/1987 [00:01<00:00, 990.01it/s]
 95%|█████████▌| 1888/1987 [00:01<00:00, 992.59it/s]
+[2026-05-30 00:40:50,862][lm_eval.evaluator][INFO] - Running loglikelihood requests
+

data/ulars/supplement/standard_eval/logs/concept-unlearning__zephyr-7b-beta_ft_lora_civil_comments_v1_ft_rmu_lora_toxic_v1__wmdp_bio.log

@@ -0,0 +1,22 @@
+repo_id: concept-unlearning/zephyr-7b-beta_ft_lora_civil_comments_v1_ft_rmu_lora_toxic_v1
+split: bio
+snapshot: /zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/concept-unlearning__zephyr-7b-beta_ft_lora_civil_comments_v1_ft_rmu_lora_toxic_v1/models--concept-unlearning--zephyr-7b-beta_ft_lora_civil_comments_v1_ft_rmu_lora_toxic_v1/snapshots/116eb7abce618baf23cadc8f0af0c1006da9adee
+cmd: /zfspool/zangyihe/conda_envs/corgo/bin/python src/eval.py experiment=eval/wmdp/default model=zephyr-7b-beta model.model_args.pretrained_model_name_or_path=/zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/concept-unlearning__zephyr-7b-beta_ft_lora_civil_comments_v1_ft_rmu_lora_toxic_v1/models--concept-unlearning--zephyr-7b-beta_ft_lora_civil_comments_v1_ft_rmu_lora_toxic_v1/snapshots/116eb7abce618baf23cadc8f0af0c1006da9adee model.tokenizer_args.pretrained_model_name_or_path=/zfspool/zangyihe/ulars/models/zephyr-7b-beta model.model_args.attn_implementation=sdpa model.model_args.torch_dtype=bfloat16 data_split=bio task_name=hf_wmdp_bio_concept-unlearning__zephyr-7b-beta_ft_lora_civil_comments_v1_ft_rmu_lora_toxic_v1 paths.output_dir=/zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/concept-unlearning__zephyr-7b-beta_ft_lora_civil_comments_v1_ft_rmu_lora_toxic_v1__wmdp_bio eval.lm_eval.overwrite=true eval.lm_eval.tasks=[wmdp_bio] eval.lm_eval.simple_evaluate_args.batch_size=16
+
+/zfspool/zangyihe/conda_envs/corgo/lib/python3.11/site-packages/transformers/utils/hub.py:105: FutureWarning: Using `TRANSFORMERS_CACHE` is deprecated and will be removed in v5 of Transformers. Use `HF_HOME` instead.
+  warnings.warn(
+[2026-05-29 23:04:53,920] [INFO] [real_accelerator.py:219:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+
+[2026-05-29 23:05:20,599][lm_eval.models.huggingface][WARNING] - `pretrained` model kwarg is not of type `str`. Many other model arguments may be ignored. Please do not launch via accelerate or use `parallelize=True` if passing an existing model this way.
+[2026-05-29 23:05:20,610][lm_eval.models.huggingface][WARNING] - Passed an already-initialized model through `pretrained`, assuming single-process call to evaluate() or custom distributed integration
+[2026-05-29 23:05:20,611][evaluator][INFO] - ***** Running LMEval evaluation suite *****
+[2026-05-29 23:05:20,611][evaluator][INFO] - Fine-grained evaluations will be saved to: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/concept-unlearning__zephyr-7b-beta_ft_lora_civil_comments_v1_ft_rmu_lora_toxic_v1__wmdp_bio/LMEval_EVAL.json
+[2026-05-29 23:05:20,612][evaluator][INFO] - Aggregated evaluations will be summarised in: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/concept-unlearning__zephyr-7b-beta_ft_lora_civil_comments_v1_ft_rmu_lora_toxic_v1__wmdp_bio/LMEval_SUMMARY.json
+[2026-05-29 23:05:20,613][lm_eval.evaluator][INFO] - Setting random seed to 0 | Setting numpy seed to 1234 | Setting torch manual seed to 1234 | Setting fewshot manual seed to 1234
+[2026-05-29 23:05:20,613][lm_eval.evaluator][INFO] - Using pre-initialized model
+[2026-05-29 23:05:28,093][lm_eval.tasks][INFO] - Selected tasks:
+[2026-05-29 23:05:28,094][lm_eval.tasks][INFO] - Task: wmdp_bio (wmdp/wmdp_bio.yaml)
+[2026-05-29 23:05:28,094][lm_eval.api.task][INFO] - Building contexts for wmdp_bio on rank 0...
+
  0%|          | 0/1273 [00:00<?, ?it/s]
  8%|▊         | 97/1273 [00:00<00:01, 964.59it/s]
 15%|█▌        | 196/1273 [00:00<00:01, 975.35it/s]
 23%|██▎       | 295/1273 [00:00<00:00, 978.84it/s]
 31%|███       | 395/1273 [00:00<00:00, 984.71it/s]
 39%|███▉      | 495/1273 [00:00<00:00, 987.69it/s]
 47%|████▋     | 595/1273 [00:00<00:00, 991.72it/s]
 55%|█████▍    | 695/1273 [00:00<00:00, 994.14it/s]
 62%|██████▏   | 795/1273 [00:00<00:00, 994.41it/s]
 70%|███████   | 895/1273 [00:00<00:00, 994.37it/s]
 78%|███████▊  | 995/1273 [00:01<00:00, 994.01it/s]
 86%|████████▌ | 1095/1273 [00:01<00:00, 993.63it/s]
 94%|█████████▍| 1195/1273 [00:01<00:00, 988.28it/s]
+[2026-05-29 23:05:29,412][lm_eval.evaluator][INFO] - Running loglikelihood requests
+

data/ulars/supplement/standard_eval/logs/concept-unlearning__zephyr-7b-beta_ft_lora_civil_comments_v1_ft_rmu_lora_toxic_v1__wmdp_cyber.log

@@ -0,0 +1,22 @@
+repo_id: concept-unlearning/zephyr-7b-beta_ft_lora_civil_comments_v1_ft_rmu_lora_toxic_v1
+split: cyber
+snapshot: /zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/concept-unlearning__zephyr-7b-beta_ft_lora_civil_comments_v1_ft_rmu_lora_toxic_v1/models--concept-unlearning--zephyr-7b-beta_ft_lora_civil_comments_v1_ft_rmu_lora_toxic_v1/snapshots/116eb7abce618baf23cadc8f0af0c1006da9adee
+cmd: /zfspool/zangyihe/conda_envs/corgo/bin/python src/eval.py experiment=eval/wmdp/default model=zephyr-7b-beta model.model_args.pretrained_model_name_or_path=/zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/concept-unlearning__zephyr-7b-beta_ft_lora_civil_comments_v1_ft_rmu_lora_toxic_v1/models--concept-unlearning--zephyr-7b-beta_ft_lora_civil_comments_v1_ft_rmu_lora_toxic_v1/snapshots/116eb7abce618baf23cadc8f0af0c1006da9adee model.tokenizer_args.pretrained_model_name_or_path=/zfspool/zangyihe/ulars/models/zephyr-7b-beta model.model_args.attn_implementation=sdpa model.model_args.torch_dtype=bfloat16 data_split=cyber task_name=hf_wmdp_cyber_concept-unlearning__zephyr-7b-beta_ft_lora_civil_comments_v1_ft_rmu_lora_toxic_v1 paths.output_dir=/zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/concept-unlearning__zephyr-7b-beta_ft_lora_civil_comments_v1_ft_rmu_lora_toxic_v1__wmdp_cyber eval.lm_eval.overwrite=true eval.lm_eval.tasks=[wmdp_cyber] eval.lm_eval.simple_evaluate_args.batch_size=16
+
+/zfspool/zangyihe/conda_envs/corgo/lib/python3.11/site-packages/transformers/utils/hub.py:105: FutureWarning: Using `TRANSFORMERS_CACHE` is deprecated and will be removed in v5 of Transformers. Use `HF_HOME` instead.
+  warnings.warn(
+[2026-05-30 00:41:08,546] [INFO] [real_accelerator.py:219:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+
+[2026-05-30 00:41:27,887][lm_eval.models.huggingface][WARNING] - `pretrained` model kwarg is not of type `str`. Many other model arguments may be ignored. Please do not launch via accelerate or use `parallelize=True` if passing an existing model this way.
+[2026-05-30 00:41:27,900][lm_eval.models.huggingface][WARNING] - Passed an already-initialized model through `pretrained`, assuming single-process call to evaluate() or custom distributed integration
+[2026-05-30 00:41:27,900][evaluator][INFO] - ***** Running LMEval evaluation suite *****
+[2026-05-30 00:41:27,901][evaluator][INFO] - Fine-grained evaluations will be saved to: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/concept-unlearning__zephyr-7b-beta_ft_lora_civil_comments_v1_ft_rmu_lora_toxic_v1__wmdp_cyber/LMEval_EVAL.json
+[2026-05-30 00:41:27,901][evaluator][INFO] - Aggregated evaluations will be summarised in: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/concept-unlearning__zephyr-7b-beta_ft_lora_civil_comments_v1_ft_rmu_lora_toxic_v1__wmdp_cyber/LMEval_SUMMARY.json
+[2026-05-30 00:41:27,902][lm_eval.evaluator][INFO] - Setting random seed to 0 | Setting numpy seed to 1234 | Setting torch manual seed to 1234 | Setting fewshot manual seed to 1234
+[2026-05-30 00:41:27,902][lm_eval.evaluator][INFO] - Using pre-initialized model
+[2026-05-30 00:41:32,005][lm_eval.tasks][INFO] - Selected tasks:
+[2026-05-30 00:41:32,006][lm_eval.tasks][INFO] - Task: wmdp_cyber (wmdp/wmdp_cyber.yaml)
+[2026-05-30 00:41:32,006][lm_eval.api.task][INFO] - Building contexts for wmdp_cyber on rank 0...
+
  0%|          | 0/1987 [00:00<?, ?it/s]
  5%|▍         | 95/1987 [00:00<00:01, 949.87it/s]
 10%|▉         | 193/1987 [00:00<00:01, 962.70it/s]
 15%|█▍        | 291/1987 [00:00<00:01, 969.63it/s]
 20%|█▉        | 390/1987 [00:00<00:01, 974.15it/s]
 25%|██▍       | 489/1987 [00:00<00:01, 977.66it/s]
 30%|██▉       | 587/1987 [00:00<00:01, 967.54it/s]
 35%|███▍      | 686/1987 [00:00<00:01, 973.84it/s]
 40%|███▉      | 785/1987 [00:00<00:01, 977.61it/s]
 44%|████▍     | 883/1987 [00:00<00:01, 973.56it/s]
 49%|████▉     | 983/1987 [00:01<00:01, 979.50it/s]
 55%|█████▍    | 1083/1987 [00:01<00:00, 984.03it/s]
 59%|█████▉    | 1182/1987 [00:01<00:00, 978.09it/s]
 64%|██████▍   | 1281/1987 [00:01<00:00, 980.98it/s]
 70%|██████▉   | 1381/1987 [00:01<00:00, 983.51it/s]
 74%|███████▍  | 1480/1987 [00:01<00:00, 984.31it/s]
 79%|███████▉  | 1579/1987 [00:01<00:00, 985.92it/s]
 84%|████████▍ | 1678/1987 [00:01<00:00, 986.61it/s]
 89%|████████▉ | 1777/1987 [00:01<00:00, 977.50it/s]
 94%|█████████▍| 1877/1987 [00:01<00:00, 981.34it/s]
 99%|█████████▉| 1977/1987 [00:02<00:00, 984.79it/s]
+[2026-05-30 00:41:34,084][lm_eval.evaluator][INFO] - Running loglikelihood requests
+

data/ulars/supplement/standard_eval/logs/concept-unlearning__zephyr-7b-beta_npo_gdr_lora_wmdp_bio_v2__wmdp_bio.log

@@ -0,0 +1,22 @@
+repo_id: concept-unlearning/zephyr-7b-beta_npo_gdr_lora_wmdp_bio_v2
+split: bio
+snapshot: /zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/concept-unlearning__zephyr-7b-beta_npo_gdr_lora_wmdp_bio_v2/models--concept-unlearning--zephyr-7b-beta_npo_gdr_lora_wmdp_bio_v2/snapshots/7c2630fc65c896c79d7b166d193c2692dde14151
+cmd: /zfspool/zangyihe/conda_envs/corgo/bin/python src/eval.py experiment=eval/wmdp/default model=zephyr-7b-beta model.model_args.pretrained_model_name_or_path=/zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/concept-unlearning__zephyr-7b-beta_npo_gdr_lora_wmdp_bio_v2/models--concept-unlearning--zephyr-7b-beta_npo_gdr_lora_wmdp_bio_v2/snapshots/7c2630fc65c896c79d7b166d193c2692dde14151 model.tokenizer_args.pretrained_model_name_or_path=/zfspool/zangyihe/ulars/models/zephyr-7b-beta model.model_args.attn_implementation=sdpa model.model_args.torch_dtype=bfloat16 data_split=bio task_name=hf_wmdp_bio_concept-unlearning__zephyr-7b-beta_npo_gdr_lora_wmdp_bio_v2 paths.output_dir=/zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/concept-unlearning__zephyr-7b-beta_npo_gdr_lora_wmdp_bio_v2__wmdp_bio eval.lm_eval.overwrite=true eval.lm_eval.tasks=[wmdp_bio] eval.lm_eval.simple_evaluate_args.batch_size=16
+
+/zfspool/zangyihe/conda_envs/corgo/lib/python3.11/site-packages/transformers/utils/hub.py:105: FutureWarning: Using `TRANSFORMERS_CACHE` is deprecated and will be removed in v5 of Transformers. Use `HF_HOME` instead.
+  warnings.warn(
+[2026-05-29 23:11:20,735] [INFO] [real_accelerator.py:219:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+
+[2026-05-29 23:11:47,223][lm_eval.models.huggingface][WARNING] - `pretrained` model kwarg is not of type `str`. Many other model arguments may be ignored. Please do not launch via accelerate or use `parallelize=True` if passing an existing model this way.
+[2026-05-29 23:11:47,237][lm_eval.models.huggingface][WARNING] - Passed an already-initialized model through `pretrained`, assuming single-process call to evaluate() or custom distributed integration
+[2026-05-29 23:11:47,238][evaluator][INFO] - ***** Running LMEval evaluation suite *****
+[2026-05-29 23:11:47,239][evaluator][INFO] - Fine-grained evaluations will be saved to: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/concept-unlearning__zephyr-7b-beta_npo_gdr_lora_wmdp_bio_v2__wmdp_bio/LMEval_EVAL.json
+[2026-05-29 23:11:47,239][evaluator][INFO] - Aggregated evaluations will be summarised in: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/concept-unlearning__zephyr-7b-beta_npo_gdr_lora_wmdp_bio_v2__wmdp_bio/LMEval_SUMMARY.json
+[2026-05-29 23:11:47,240][lm_eval.evaluator][INFO] - Setting random seed to 0 | Setting numpy seed to 1234 | Setting torch manual seed to 1234 | Setting fewshot manual seed to 1234
+[2026-05-29 23:11:47,240][lm_eval.evaluator][INFO] - Using pre-initialized model
+[2026-05-29 23:11:53,821][lm_eval.tasks][INFO] - Selected tasks:
+[2026-05-29 23:11:53,821][lm_eval.tasks][INFO] - Task: wmdp_bio (wmdp/wmdp_bio.yaml)
+[2026-05-29 23:11:53,821][lm_eval.api.task][INFO] - Building contexts for wmdp_bio on rank 0...
+
  0%|          | 0/1273 [00:00<?, ?it/s]
  8%|▊         | 97/1273 [00:00<00:01, 962.04it/s]
 15%|█▌        | 195/1273 [00:00<00:01, 971.17it/s]
 23%|██▎       | 294/1273 [00:00<00:01, 977.46it/s]
 31%|███       | 393/1273 [00:00<00:00, 981.77it/s]
 39%|███▊      | 492/1273 [00:00<00:00, 984.53it/s]
 46%|████▋     | 591/1273 [00:00<00:00, 985.53it/s]
 54%|█████▍    | 691/1273 [00:00<00:00, 987.77it/s]
 62%|██████▏   | 791/1273 [00:00<00:00, 988.94it/s]
 70%|██████▉   | 891/1273 [00:00<00:00, 989.35it/s]
 78%|███████▊  | 991/1273 [00:01<00:00, 991.40it/s]
 86%|████████▌ | 1091/1273 [00:01<00:00, 993.32it/s]
 94%|█████████▎| 1191/1273 [00:01<00:00, 994.94it/s]
+[2026-05-29 23:11:55,141][lm_eval.evaluator][INFO] - Running loglikelihood requests
+

data/ulars/supplement/standard_eval/logs/concept-unlearning__zephyr-7b-beta_npo_gdr_lora_wmdp_bio_v2__wmdp_cyber.log

@@ -0,0 +1,22 @@
+repo_id: concept-unlearning/zephyr-7b-beta_npo_gdr_lora_wmdp_bio_v2
+split: cyber
+snapshot: /zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/concept-unlearning__zephyr-7b-beta_npo_gdr_lora_wmdp_bio_v2/models--concept-unlearning--zephyr-7b-beta_npo_gdr_lora_wmdp_bio_v2/snapshots/7c2630fc65c896c79d7b166d193c2692dde14151
+cmd: /zfspool/zangyihe/conda_envs/corgo/bin/python src/eval.py experiment=eval/wmdp/default model=zephyr-7b-beta model.model_args.pretrained_model_name_or_path=/zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/concept-unlearning__zephyr-7b-beta_npo_gdr_lora_wmdp_bio_v2/models--concept-unlearning--zephyr-7b-beta_npo_gdr_lora_wmdp_bio_v2/snapshots/7c2630fc65c896c79d7b166d193c2692dde14151 model.tokenizer_args.pretrained_model_name_or_path=/zfspool/zangyihe/ulars/models/zephyr-7b-beta model.model_args.attn_implementation=sdpa model.model_args.torch_dtype=bfloat16 data_split=cyber task_name=hf_wmdp_cyber_concept-unlearning__zephyr-7b-beta_npo_gdr_lora_wmdp_bio_v2 paths.output_dir=/zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/concept-unlearning__zephyr-7b-beta_npo_gdr_lora_wmdp_bio_v2__wmdp_cyber eval.lm_eval.overwrite=true eval.lm_eval.tasks=[wmdp_cyber] eval.lm_eval.simple_evaluate_args.batch_size=16
+
+/zfspool/zangyihe/conda_envs/corgo/lib/python3.11/site-packages/transformers/utils/hub.py:105: FutureWarning: Using `TRANSFORMERS_CACHE` is deprecated and will be removed in v5 of Transformers. Use `HF_HOME` instead.
+  warnings.warn(
+[2026-05-30 00:42:53,169] [INFO] [real_accelerator.py:219:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+
+[2026-05-30 00:43:00,830][lm_eval.models.huggingface][WARNING] - `pretrained` model kwarg is not of type `str`. Many other model arguments may be ignored. Please do not launch via accelerate or use `parallelize=True` if passing an existing model this way.
+[2026-05-30 00:43:00,842][lm_eval.models.huggingface][WARNING] - Passed an already-initialized model through `pretrained`, assuming single-process call to evaluate() or custom distributed integration
+[2026-05-30 00:43:00,843][evaluator][INFO] - ***** Running LMEval evaluation suite *****
+[2026-05-30 00:43:00,843][evaluator][INFO] - Fine-grained evaluations will be saved to: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/concept-unlearning__zephyr-7b-beta_npo_gdr_lora_wmdp_bio_v2__wmdp_cyber/LMEval_EVAL.json
+[2026-05-30 00:43:00,843][evaluator][INFO] - Aggregated evaluations will be summarised in: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/concept-unlearning__zephyr-7b-beta_npo_gdr_lora_wmdp_bio_v2__wmdp_cyber/LMEval_SUMMARY.json
+[2026-05-30 00:43:00,844][lm_eval.evaluator][INFO] - Setting random seed to 0 | Setting numpy seed to 1234 | Setting torch manual seed to 1234 | Setting fewshot manual seed to 1234
+[2026-05-30 00:43:00,844][lm_eval.evaluator][INFO] - Using pre-initialized model
+[2026-05-30 00:43:06,738][lm_eval.tasks][INFO] - Selected tasks:
+[2026-05-30 00:43:06,738][lm_eval.tasks][INFO] - Task: wmdp_cyber (wmdp/wmdp_cyber.yaml)
+[2026-05-30 00:43:06,738][lm_eval.api.task][INFO] - Building contexts for wmdp_cyber on rank 0...
+
  0%|          | 0/1987 [00:00<?, ?it/s]
  5%|▍         | 95/1987 [00:00<00:02, 944.58it/s]
 10%|▉         | 194/1987 [00:00<00:01, 968.93it/s]
 15%|█▍        | 293/1987 [00:00<00:01, 978.38it/s]
 20%|█▉        | 393/1987 [00:00<00:01, 983.06it/s]
 25%|██▍       | 493/1987 [00:00<00:01, 988.75it/s]
 30%|██▉       | 592/1987 [00:00<00:01, 978.91it/s]
 35%|███▍      | 692/1987 [00:00<00:01, 983.69it/s]
 40%|███▉      | 791/1987 [00:00<00:01, 981.84it/s]
 45%|████▍     | 891/1987 [00:00<00:01, 985.50it/s]
 50%|████▉     | 991/1987 [00:01<00:01, 989.33it/s]
 55%|█████▍    | 1092/1987 [00:01<00:00, 993.28it/s]
 60%|██████    | 1193/1987 [00:01<00:00, 995.40it/s]
 65%|██████▌   | 1293/1987 [00:01<00:00, 990.66it/s]
 70%|███████   | 1393/1987 [00:01<00:00, 992.26it/s]
 75%|███████▌  | 1494/1987 [00:01<00:00, 996.32it/s]
 80%|████████  | 1595/1987 [00:01<00:00, 998.74it/s]
 85%|████████▌ | 1696/1987 [00:01<00:00, 999.75it/s]
 90%|█████████ | 1797/1987 [00:01<00:00, 1000.17it/s]
 96%|█████████▌| 1898/1987 [00:01<00:00, 996.94it/s] 
+[2026-05-30 00:43:08,794][lm_eval.evaluator][INFO] - Running loglikelihood requests
+

data/ulars/supplement/standard_eval/logs/flyingbugs__WMDP_GradDiff_WAGLE_Zephyr_7B__wmdp_bio.log

@@ -0,0 +1,22 @@
+repo_id: flyingbugs/WMDP_GradDiff_WAGLE_Zephyr_7B
+split: bio
+snapshot: /zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/flyingbugs__WMDP_GradDiff_WAGLE_Zephyr_7B/models--flyingbugs--WMDP_GradDiff_WAGLE_Zephyr_7B/snapshots/71be9e8e0a83534401a7d328c959e75ebdf66b71
+cmd: /zfspool/zangyihe/conda_envs/corgo/bin/python src/eval.py experiment=eval/wmdp/default model=zephyr-7b-beta model.model_args.pretrained_model_name_or_path=/zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/flyingbugs__WMDP_GradDiff_WAGLE_Zephyr_7B/models--flyingbugs--WMDP_GradDiff_WAGLE_Zephyr_7B/snapshots/71be9e8e0a83534401a7d328c959e75ebdf66b71 model.tokenizer_args.pretrained_model_name_or_path=/zfspool/zangyihe/ulars/models/zephyr-7b-beta model.model_args.attn_implementation=sdpa model.model_args.torch_dtype=bfloat16 data_split=bio task_name=hf_wmdp_bio_flyingbugs__WMDP_GradDiff_WAGLE_Zephyr_7B paths.output_dir=/zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/flyingbugs__WMDP_GradDiff_WAGLE_Zephyr_7B__wmdp_bio eval.lm_eval.overwrite=true eval.lm_eval.tasks=[wmdp_bio] eval.lm_eval.simple_evaluate_args.batch_size=16
+
+/zfspool/zangyihe/conda_envs/corgo/lib/python3.11/site-packages/transformers/utils/hub.py:105: FutureWarning: Using `TRANSFORMERS_CACHE` is deprecated and will be removed in v5 of Transformers. Use `HF_HOME` instead.
+  warnings.warn(
+[2026-05-29 21:58:17,456] [INFO] [real_accelerator.py:219:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+
+[2026-05-29 21:58:59,863][lm_eval.models.huggingface][WARNING] - `pretrained` model kwarg is not of type `str`. Many other model arguments may be ignored. Please do not launch via accelerate or use `parallelize=True` if passing an existing model this way.
+[2026-05-29 21:58:59,877][lm_eval.models.huggingface][WARNING] - Passed an already-initialized model through `pretrained`, assuming single-process call to evaluate() or custom distributed integration
+[2026-05-29 21:58:59,878][evaluator][INFO] - ***** Running LMEval evaluation suite *****
+[2026-05-29 21:58:59,878][evaluator][INFO] - Fine-grained evaluations will be saved to: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/flyingbugs__WMDP_GradDiff_WAGLE_Zephyr_7B__wmdp_bio/LMEval_EVAL.json
+[2026-05-29 21:58:59,878][evaluator][INFO] - Aggregated evaluations will be summarised in: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/flyingbugs__WMDP_GradDiff_WAGLE_Zephyr_7B__wmdp_bio/LMEval_SUMMARY.json
+[2026-05-29 21:58:59,879][lm_eval.evaluator][INFO] - Setting random seed to 0 | Setting numpy seed to 1234 | Setting torch manual seed to 1234 | Setting fewshot manual seed to 1234
+[2026-05-29 21:58:59,880][lm_eval.evaluator][INFO] - Using pre-initialized model
+[2026-05-29 21:59:06,648][lm_eval.tasks][INFO] - Selected tasks:
+[2026-05-29 21:59:06,648][lm_eval.tasks][INFO] - Task: wmdp_bio (wmdp/wmdp_bio.yaml)
+[2026-05-29 21:59:06,648][lm_eval.api.task][INFO] - Building contexts for wmdp_bio on rank 0...
+
  0%|          | 0/1273 [00:00<?, ?it/s]
  7%|▋         | 88/1273 [00:00<00:01, 872.20it/s]
 15%|█▍        | 186/1273 [00:00<00:01, 933.78it/s]
 22%|██▏       | 280/1273 [00:00<00:01, 917.76it/s]
 29%|██▉       | 372/1273 [00:00<00:01, 898.74it/s]
 37%|███▋      | 465/1273 [00:00<00:00, 906.67it/s]
 44%|████▎     | 556/1273 [00:00<00:00, 893.92it/s]
 51%|█████     | 646/1273 [00:00<00:00, 888.86it/s]
 58%|█████▊    | 736/1273 [00:00<00:00, 889.55it/s]
 65%|██████▍   | 825/1273 [00:00<00:00, 884.21it/s]
 72%|███████▏  | 914/1273 [00:01<00:00, 878.40it/s]
 79%|███████▉  | 1012/1273 [00:01<00:00, 907.25it/s]
 87%|████████▋ | 1111/1273 [00:01<00:00, 930.90it/s]
 95%|█████████▍| 1205/1273 [00:01<00:00, 920.91it/s]
+[2026-05-29 21:59:08,080][lm_eval.evaluator][INFO] - Running loglikelihood requests
+

data/ulars/supplement/standard_eval/logs/flyingbugs__WMDP_GradDiff_WAGLE_Zephyr_7B__wmdp_cyber.log

@@ -0,0 +1,22 @@
+repo_id: flyingbugs/WMDP_GradDiff_WAGLE_Zephyr_7B
+split: cyber
+snapshot: /zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/flyingbugs__WMDP_GradDiff_WAGLE_Zephyr_7B/models--flyingbugs--WMDP_GradDiff_WAGLE_Zephyr_7B/snapshots/71be9e8e0a83534401a7d328c959e75ebdf66b71
+cmd: /zfspool/zangyihe/conda_envs/corgo/bin/python src/eval.py experiment=eval/wmdp/default model=zephyr-7b-beta model.model_args.pretrained_model_name_or_path=/zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/flyingbugs__WMDP_GradDiff_WAGLE_Zephyr_7B/models--flyingbugs--WMDP_GradDiff_WAGLE_Zephyr_7B/snapshots/71be9e8e0a83534401a7d328c959e75ebdf66b71 model.tokenizer_args.pretrained_model_name_or_path=/zfspool/zangyihe/ulars/models/zephyr-7b-beta model.model_args.attn_implementation=sdpa model.model_args.torch_dtype=bfloat16 data_split=cyber task_name=hf_wmdp_cyber_flyingbugs__WMDP_GradDiff_WAGLE_Zephyr_7B paths.output_dir=/zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/flyingbugs__WMDP_GradDiff_WAGLE_Zephyr_7B__wmdp_cyber eval.lm_eval.overwrite=true eval.lm_eval.tasks=[wmdp_cyber] eval.lm_eval.simple_evaluate_args.batch_size=16
+
+/zfspool/zangyihe/conda_envs/corgo/lib/python3.11/site-packages/transformers/utils/hub.py:105: FutureWarning: Using `TRANSFORMERS_CACHE` is deprecated and will be removed in v5 of Transformers. Use `HF_HOME` instead.
+  warnings.warn(
+[2026-05-30 00:13:41,364] [INFO] [real_accelerator.py:219:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+
+[2026-05-30 00:14:37,687][lm_eval.models.huggingface][WARNING] - `pretrained` model kwarg is not of type `str`. Many other model arguments may be ignored. Please do not launch via accelerate or use `parallelize=True` if passing an existing model this way.
+[2026-05-30 00:14:37,700][lm_eval.models.huggingface][WARNING] - Passed an already-initialized model through `pretrained`, assuming single-process call to evaluate() or custom distributed integration
+[2026-05-30 00:14:37,700][evaluator][INFO] - ***** Running LMEval evaluation suite *****
+[2026-05-30 00:14:37,701][evaluator][INFO] - Fine-grained evaluations will be saved to: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/flyingbugs__WMDP_GradDiff_WAGLE_Zephyr_7B__wmdp_cyber/LMEval_EVAL.json
+[2026-05-30 00:14:37,701][evaluator][INFO] - Aggregated evaluations will be summarised in: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/flyingbugs__WMDP_GradDiff_WAGLE_Zephyr_7B__wmdp_cyber/LMEval_SUMMARY.json
+[2026-05-30 00:14:37,702][lm_eval.evaluator][INFO] - Setting random seed to 0 | Setting numpy seed to 1234 | Setting torch manual seed to 1234 | Setting fewshot manual seed to 1234
+[2026-05-30 00:14:37,702][lm_eval.evaluator][INFO] - Using pre-initialized model
+[2026-05-30 00:14:42,934][lm_eval.tasks][INFO] - Selected tasks:
+[2026-05-30 00:14:42,935][lm_eval.tasks][INFO] - Task: wmdp_cyber (wmdp/wmdp_cyber.yaml)
+[2026-05-30 00:14:42,935][lm_eval.api.task][INFO] - Building contexts for wmdp_cyber on rank 0...
+
  0%|          | 0/1987 [00:00<?, ?it/s]
  5%|▍         | 97/1987 [00:00<00:01, 969.31it/s]
 10%|▉         | 197/1987 [00:00<00:01, 981.73it/s]
 15%|█▍        | 297/1987 [00:00<00:01, 986.53it/s]
 20%|█▉        | 397/1987 [00:00<00:01, 991.49it/s]
 25%|██▌       | 497/1987 [00:00<00:01, 993.97it/s]
 30%|███       | 598/1987 [00:00<00:01, 997.80it/s]
 35%|███▌      | 699/1987 [00:00<00:01, 1000.39it/s]
 40%|████      | 800/1987 [00:00<00:01, 1001.11it/s]
 45%|████▌     | 901/1987 [00:00<00:01, 1001.79it/s]
 50%|█████     | 1002/1987 [00:01<00:00, 1004.11it/s]
 56%|█████▌    | 1103/1987 [00:01<00:00, 1005.14it/s]
 61%|██████    | 1204/1987 [00:01<00:00, 995.44it/s] 
 66%|██████▌   | 1305/1987 [00:01<00:00, 999.13it/s]
 71%|███████   | 1406/1987 [00:01<00:00, 1000.41it/s]
 76%|███████▌  | 1507/1987 [00:01<00:00, 1001.96it/s]
 81%|████████  | 1608/1987 [00:01<00:00, 1002.07it/s]
 86%|████████▌ | 1709/1987 [00:01<00:00, 1003.35it/s]
 91%|█████████ | 1810/1987 [00:01<00:00, 1005.33it/s]
 96%|█████████▌| 1911/1987 [00:01<00:00, 1006.17it/s]
+[2026-05-30 00:14:44,970][lm_eval.evaluator][INFO] - Running loglikelihood requests
+

data/ulars/supplement/standard_eval/logs/flyingbugs__WMDP_NPO_WAGLE_Zephyr_7B__wmdp_bio.log

@@ -0,0 +1,22 @@
+repo_id: flyingbugs/WMDP_NPO_WAGLE_Zephyr_7B
+split: bio
+snapshot: /zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/flyingbugs__WMDP_NPO_WAGLE_Zephyr_7B/models--flyingbugs--WMDP_NPO_WAGLE_Zephyr_7B/snapshots/a432fce44c7715272d9aab47e91f5775a4cb7a82
+cmd: /zfspool/zangyihe/conda_envs/corgo/bin/python src/eval.py experiment=eval/wmdp/default model=zephyr-7b-beta model.model_args.pretrained_model_name_or_path=/zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/flyingbugs__WMDP_NPO_WAGLE_Zephyr_7B/models--flyingbugs--WMDP_NPO_WAGLE_Zephyr_7B/snapshots/a432fce44c7715272d9aab47e91f5775a4cb7a82 model.tokenizer_args.pretrained_model_name_or_path=/zfspool/zangyihe/ulars/models/zephyr-7b-beta model.model_args.attn_implementation=sdpa model.model_args.torch_dtype=bfloat16 data_split=bio task_name=hf_wmdp_bio_flyingbugs__WMDP_NPO_WAGLE_Zephyr_7B paths.output_dir=/zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/flyingbugs__WMDP_NPO_WAGLE_Zephyr_7B__wmdp_bio eval.lm_eval.overwrite=true eval.lm_eval.tasks=[wmdp_bio] eval.lm_eval.simple_evaluate_args.batch_size=16
+
+/zfspool/zangyihe/conda_envs/corgo/lib/python3.11/site-packages/transformers/utils/hub.py:105: FutureWarning: Using `TRANSFORMERS_CACHE` is deprecated and will be removed in v5 of Transformers. Use `HF_HOME` instead.
+  warnings.warn(
+[2026-05-29 21:58:15,599] [INFO] [real_accelerator.py:219:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+
+[2026-05-29 21:58:56,964][lm_eval.models.huggingface][WARNING] - `pretrained` model kwarg is not of type `str`. Many other model arguments may be ignored. Please do not launch via accelerate or use `parallelize=True` if passing an existing model this way.
+[2026-05-29 21:58:56,976][lm_eval.models.huggingface][WARNING] - Passed an already-initialized model through `pretrained`, assuming single-process call to evaluate() or custom distributed integration
+[2026-05-29 21:58:56,976][evaluator][INFO] - ***** Running LMEval evaluation suite *****
+[2026-05-29 21:58:56,977][evaluator][INFO] - Fine-grained evaluations will be saved to: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/flyingbugs__WMDP_NPO_WAGLE_Zephyr_7B__wmdp_bio/LMEval_EVAL.json
+[2026-05-29 21:58:56,977][evaluator][INFO] - Aggregated evaluations will be summarised in: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/flyingbugs__WMDP_NPO_WAGLE_Zephyr_7B__wmdp_bio/LMEval_SUMMARY.json
+[2026-05-29 21:58:56,978][lm_eval.evaluator][INFO] - Setting random seed to 0 | Setting numpy seed to 1234 | Setting torch manual seed to 1234 | Setting fewshot manual seed to 1234
+[2026-05-29 21:58:56,978][lm_eval.evaluator][INFO] - Using pre-initialized model
+[2026-05-29 21:59:08,282][lm_eval.tasks][INFO] - Selected tasks:
+[2026-05-29 21:59:08,283][lm_eval.tasks][INFO] - Task: wmdp_bio (wmdp/wmdp_bio.yaml)
+[2026-05-29 21:59:08,283][lm_eval.api.task][INFO] - Building contexts for wmdp_bio on rank 0...
+
  0%|          | 0/1273 [00:00<?, ?it/s]
  7%|▋         | 90/1273 [00:00<00:01, 892.07it/s]
 14%|█▍        | 180/1273 [00:00<00:01, 882.72it/s]
 22%|██▏       | 276/1273 [00:00<00:01, 917.10it/s]
 29%|██▉       | 372/1273 [00:00<00:00, 933.90it/s]
 37%|███▋      | 470/1273 [00:00<00:00, 947.70it/s]
 45%|████▍     | 568/1273 [00:00<00:00, 957.54it/s]
 52%|█████▏    | 667/1273 [00:00<00:00, 966.60it/s]
 60%|██████    | 765/1273 [00:00<00:00, 969.08it/s]
 68%|██████▊   | 864/1273 [00:00<00:00, 974.35it/s]
 76%|███████▌  | 962/1273 [00:01<00:00, 970.85it/s]
 83%|████████▎ | 1060/1273 [00:01<00:00, 972.60it/s]
 91%|█████████ | 1158/1273 [00:01<00:00, 970.61it/s]
 99%|█████████▊| 1257/1273 [00:01<00:00, 975.51it/s]
+[2026-05-29 21:59:09,647][lm_eval.evaluator][INFO] - Running loglikelihood requests
+

data/ulars/supplement/standard_eval/logs/flyingbugs__WMDP_NPO_WAGLE_Zephyr_7B__wmdp_cyber.log

@@ -0,0 +1,22 @@
+repo_id: flyingbugs/WMDP_NPO_WAGLE_Zephyr_7B
+split: cyber
+snapshot: /zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/flyingbugs__WMDP_NPO_WAGLE_Zephyr_7B/models--flyingbugs--WMDP_NPO_WAGLE_Zephyr_7B/snapshots/a432fce44c7715272d9aab47e91f5775a4cb7a82
+cmd: /zfspool/zangyihe/conda_envs/corgo/bin/python src/eval.py experiment=eval/wmdp/default model=zephyr-7b-beta model.model_args.pretrained_model_name_or_path=/zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/flyingbugs__WMDP_NPO_WAGLE_Zephyr_7B/models--flyingbugs--WMDP_NPO_WAGLE_Zephyr_7B/snapshots/a432fce44c7715272d9aab47e91f5775a4cb7a82 model.tokenizer_args.pretrained_model_name_or_path=/zfspool/zangyihe/ulars/models/zephyr-7b-beta model.model_args.attn_implementation=sdpa model.model_args.torch_dtype=bfloat16 data_split=cyber task_name=hf_wmdp_cyber_flyingbugs__WMDP_NPO_WAGLE_Zephyr_7B paths.output_dir=/zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/flyingbugs__WMDP_NPO_WAGLE_Zephyr_7B__wmdp_cyber eval.lm_eval.overwrite=true eval.lm_eval.tasks=[wmdp_cyber] eval.lm_eval.simple_evaluate_args.batch_size=16
+
+/zfspool/zangyihe/conda_envs/corgo/lib/python3.11/site-packages/transformers/utils/hub.py:105: FutureWarning: Using `TRANSFORMERS_CACHE` is deprecated and will be removed in v5 of Transformers. Use `HF_HOME` instead.
+  warnings.warn(
+[2026-05-30 00:15:11,464] [INFO] [real_accelerator.py:219:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+
+[2026-05-30 00:15:25,978][lm_eval.models.huggingface][WARNING] - `pretrained` model kwarg is not of type `str`. Many other model arguments may be ignored. Please do not launch via accelerate or use `parallelize=True` if passing an existing model this way.
+[2026-05-30 00:15:25,991][lm_eval.models.huggingface][WARNING] - Passed an already-initialized model through `pretrained`, assuming single-process call to evaluate() or custom distributed integration
+[2026-05-30 00:15:25,992][evaluator][INFO] - ***** Running LMEval evaluation suite *****
+[2026-05-30 00:15:25,992][evaluator][INFO] - Fine-grained evaluations will be saved to: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/flyingbugs__WMDP_NPO_WAGLE_Zephyr_7B__wmdp_cyber/LMEval_EVAL.json
+[2026-05-30 00:15:25,992][evaluator][INFO] - Aggregated evaluations will be summarised in: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/flyingbugs__WMDP_NPO_WAGLE_Zephyr_7B__wmdp_cyber/LMEval_SUMMARY.json
+[2026-05-30 00:15:25,993][lm_eval.evaluator][INFO] - Setting random seed to 0 | Setting numpy seed to 1234 | Setting torch manual seed to 1234 | Setting fewshot manual seed to 1234
+[2026-05-30 00:15:25,994][lm_eval.evaluator][INFO] - Using pre-initialized model
+[2026-05-30 00:15:34,566][lm_eval.tasks][INFO] - Selected tasks:
+[2026-05-30 00:15:34,567][lm_eval.tasks][INFO] - Task: wmdp_cyber (wmdp/wmdp_cyber.yaml)
+[2026-05-30 00:15:34,567][lm_eval.api.task][INFO] - Building contexts for wmdp_cyber on rank 0...
+
  0%|          | 0/1987 [00:00<?, ?it/s]
  5%|▍         | 96/1987 [00:00<00:01, 957.04it/s]
 10%|▉         | 194/1987 [00:00<00:01, 966.44it/s]
 15%|█▍        | 293/1987 [00:00<00:01, 973.65it/s]
 20%|█▉        | 392/1987 [00:00<00:01, 978.38it/s]
 25%|██▍       | 490/1987 [00:00<00:01, 977.63it/s]
 30%|██▉       | 589/1987 [00:00<00:01, 979.28it/s]
 35%|███▍      | 688/1987 [00:00<00:01, 982.32it/s]
 40%|███▉      | 787/1987 [00:00<00:01, 983.52it/s]
 45%|████▍     | 886/1987 [00:00<00:01, 985.22it/s]
 50%|████▉     | 985/1987 [00:01<00:01, 986.52it/s]
 55%|█████▍    | 1084/1987 [00:01<00:00, 986.64it/s]
 60%|█████▉    | 1184/1987 [00:01<00:00, 989.25it/s]
 65%|██████▍   | 1283/1987 [00:01<00:00, 989.30it/s]
 70%|██████▉   | 1383/1987 [00:01<00:00, 990.90it/s]
 75%|███████▍  | 1483/1987 [00:01<00:00, 993.14it/s]
 80%|███████▉  | 1583/1987 [00:01<00:00, 994.90it/s]
 85%|████████▍ | 1683/1987 [00:01<00:00, 995.70it/s]
 90%|████████▉ | 1783/1987 [00:01<00:00, 988.15it/s]
 95%|█████████▍| 1883/1987 [00:01<00:00, 990.39it/s]
+[2026-05-30 00:15:36,630][lm_eval.evaluator][INFO] - Running loglikelihood requests
+

data/ulars/supplement/standard_eval/logs/flyingbugs__bi_unlearn_wmdp__wmdp_bio.log

@@ -0,0 +1,22 @@
+repo_id: flyingbugs/bi_unlearn_wmdp
+split: bio
+snapshot: /zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/flyingbugs__bi_unlearn_wmdp/models--flyingbugs--bi_unlearn_wmdp/snapshots/6bf85c36d2dda6271668df3eb7628c1e4e2cae60
+cmd: /zfspool/zangyihe/conda_envs/corgo/bin/python src/eval.py experiment=eval/wmdp/default model=zephyr-7b-beta model.model_args.pretrained_model_name_or_path=/zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/flyingbugs__bi_unlearn_wmdp/models--flyingbugs--bi_unlearn_wmdp/snapshots/6bf85c36d2dda6271668df3eb7628c1e4e2cae60 model.tokenizer_args.pretrained_model_name_or_path=/zfspool/zangyihe/ulars/models/zephyr-7b-beta model.model_args.attn_implementation=sdpa model.model_args.torch_dtype=bfloat16 data_split=bio task_name=hf_wmdp_bio_flyingbugs__bi_unlearn_wmdp paths.output_dir=/zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/flyingbugs__bi_unlearn_wmdp__wmdp_bio eval.lm_eval.overwrite=true eval.lm_eval.tasks=[wmdp_bio] eval.lm_eval.simple_evaluate_args.batch_size=16
+
+/zfspool/zangyihe/conda_envs/corgo/lib/python3.11/site-packages/transformers/utils/hub.py:105: FutureWarning: Using `TRANSFORMERS_CACHE` is deprecated and will be removed in v5 of Transformers. Use `HF_HOME` instead.
+  warnings.warn(
+[2026-05-29 22:18:25,026] [INFO] [real_accelerator.py:219:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+
+[2026-05-29 22:18:57,852][lm_eval.models.huggingface][WARNING] - `pretrained` model kwarg is not of type `str`. Many other model arguments may be ignored. Please do not launch via accelerate or use `parallelize=True` if passing an existing model this way.
+[2026-05-29 22:18:57,864][lm_eval.models.huggingface][WARNING] - Passed an already-initialized model through `pretrained`, assuming single-process call to evaluate() or custom distributed integration
+[2026-05-29 22:18:57,864][evaluator][INFO] - ***** Running LMEval evaluation suite *****
+[2026-05-29 22:18:57,865][evaluator][INFO] - Fine-grained evaluations will be saved to: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/flyingbugs__bi_unlearn_wmdp__wmdp_bio/LMEval_EVAL.json
+[2026-05-29 22:18:57,865][evaluator][INFO] - Aggregated evaluations will be summarised in: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/flyingbugs__bi_unlearn_wmdp__wmdp_bio/LMEval_SUMMARY.json
+[2026-05-29 22:18:57,866][lm_eval.evaluator][INFO] - Setting random seed to 0 | Setting numpy seed to 1234 | Setting torch manual seed to 1234 | Setting fewshot manual seed to 1234
+[2026-05-29 22:18:57,866][lm_eval.evaluator][INFO] - Using pre-initialized model
+[2026-05-29 22:19:12,488][lm_eval.tasks][INFO] - Selected tasks:
+[2026-05-29 22:19:12,488][lm_eval.tasks][INFO] - Task: wmdp_bio (wmdp/wmdp_bio.yaml)
+[2026-05-29 22:19:12,489][lm_eval.api.task][INFO] - Building contexts for wmdp_bio on rank 0...
+
  0%|          | 0/1273 [00:00<?, ?it/s]
  7%|▋         | 94/1273 [00:00<00:01, 930.20it/s]
 15%|█▍        | 188/1273 [00:00<00:01, 932.77it/s]
 22%|██▏       | 282/1273 [00:00<00:01, 933.26it/s]
 30%|██▉       | 376/1273 [00:00<00:00, 930.32it/s]
 37%|███▋      | 471/1273 [00:00<00:00, 934.65it/s]
 44%|████▍     | 566/1273 [00:00<00:00, 936.61it/s]
 52%|█████▏    | 661/1273 [00:00<00:00, 940.78it/s]
 59%|█████▉    | 756/1273 [00:00<00:00, 942.96it/s]
 67%|██████▋   | 852/1273 [00:00<00:00, 945.32it/s]
 74%|███████▍  | 947/1273 [00:01<00:00, 941.57it/s]
 82%|████████▏ | 1043/1273 [00:01<00:00, 946.82it/s]
 89%|████████▉ | 1139/1273 [00:01<00:00, 950.43it/s]
 97%|█████████▋| 1235/1273 [00:01<00:00, 950.83it/s]
+[2026-05-29 22:19:13,870][lm_eval.evaluator][INFO] - Running loglikelihood requests
+

data/ulars/supplement/standard_eval/logs/flyingbugs__bi_unlearn_wmdp__wmdp_cyber.log

@@ -0,0 +1,22 @@
+repo_id: flyingbugs/bi_unlearn_wmdp
+split: cyber
+snapshot: /zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/flyingbugs__bi_unlearn_wmdp/models--flyingbugs--bi_unlearn_wmdp/snapshots/6bf85c36d2dda6271668df3eb7628c1e4e2cae60
+cmd: /zfspool/zangyihe/conda_envs/corgo/bin/python src/eval.py experiment=eval/wmdp/default model=zephyr-7b-beta model.model_args.pretrained_model_name_or_path=/zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/flyingbugs__bi_unlearn_wmdp/models--flyingbugs--bi_unlearn_wmdp/snapshots/6bf85c36d2dda6271668df3eb7628c1e4e2cae60 model.tokenizer_args.pretrained_model_name_or_path=/zfspool/zangyihe/ulars/models/zephyr-7b-beta model.model_args.attn_implementation=sdpa model.model_args.torch_dtype=bfloat16 data_split=cyber task_name=hf_wmdp_cyber_flyingbugs__bi_unlearn_wmdp paths.output_dir=/zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/flyingbugs__bi_unlearn_wmdp__wmdp_cyber eval.lm_eval.overwrite=true eval.lm_eval.tasks=[wmdp_cyber] eval.lm_eval.simple_evaluate_args.batch_size=16
+
+/zfspool/zangyihe/conda_envs/corgo/lib/python3.11/site-packages/transformers/utils/hub.py:105: FutureWarning: Using `TRANSFORMERS_CACHE` is deprecated and will be removed in v5 of Transformers. Use `HF_HOME` instead.
+  warnings.warn(
+[2026-05-30 00:16:46,228] [INFO] [real_accelerator.py:219:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+
+[2026-05-30 00:17:02,969][lm_eval.models.huggingface][WARNING] - `pretrained` model kwarg is not of type `str`. Many other model arguments may be ignored. Please do not launch via accelerate or use `parallelize=True` if passing an existing model this way.
+[2026-05-30 00:17:02,981][lm_eval.models.huggingface][WARNING] - Passed an already-initialized model through `pretrained`, assuming single-process call to evaluate() or custom distributed integration
+[2026-05-30 00:17:02,982][evaluator][INFO] - ***** Running LMEval evaluation suite *****
+[2026-05-30 00:17:02,982][evaluator][INFO] - Fine-grained evaluations will be saved to: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/flyingbugs__bi_unlearn_wmdp__wmdp_cyber/LMEval_EVAL.json
+[2026-05-30 00:17:02,983][evaluator][INFO] - Aggregated evaluations will be summarised in: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/flyingbugs__bi_unlearn_wmdp__wmdp_cyber/LMEval_SUMMARY.json
+[2026-05-30 00:17:02,984][lm_eval.evaluator][INFO] - Setting random seed to 0 | Setting numpy seed to 1234 | Setting torch manual seed to 1234 | Setting fewshot manual seed to 1234
+[2026-05-30 00:17:02,984][lm_eval.evaluator][INFO] - Using pre-initialized model
+[2026-05-30 00:17:12,687][lm_eval.tasks][INFO] - Selected tasks:
+[2026-05-30 00:17:12,687][lm_eval.tasks][INFO] - Task: wmdp_cyber (wmdp/wmdp_cyber.yaml)
+[2026-05-30 00:17:12,688][lm_eval.api.task][INFO] - Building contexts for wmdp_cyber on rank 0...
+
  0%|          | 0/1987 [00:00<?, ?it/s]
  5%|▍         | 96/1987 [00:00<00:01, 954.51it/s]
 10%|▉         | 194/1987 [00:00<00:01, 968.49it/s]
 15%|█▍        | 291/1987 [00:00<00:01, 956.01it/s]
 20%|█▉        | 389/1987 [00:00<00:01, 964.95it/s]
 25%|██▍       | 487/1987 [00:00<00:01, 968.97it/s]
 29%|██▉       | 586/1987 [00:00<00:01, 973.23it/s]
 34%|███▍      | 685/1987 [00:00<00:01, 976.83it/s]
 40%|███▉      | 785/1987 [00:00<00:01, 981.57it/s]
 45%|████▍     | 885/1987 [00:00<00:01, 986.24it/s]
 50%|████▉     | 985/1987 [00:01<00:01, 989.97it/s]
 55%|█████▍    | 1085/1987 [00:01<00:00, 988.55it/s]
 60%|█████▉    | 1185/1987 [00:01<00:00, 991.81it/s]
 65%|██████▍   | 1285/1987 [00:01<00:00, 991.07it/s]
 70%|██████▉   | 1386/1987 [00:01<00:00, 993.84it/s]
 75%|███████▍  | 1486/1987 [00:01<00:00, 995.54it/s]
 80%|███████▉  | 1586/1987 [00:01<00:00, 995.99it/s]
 85%|████████▍ | 1687/1987 [00:01<00:00, 997.92it/s]
 90%|████████▉ | 1787/1987 [00:01<00:00, 996.39it/s]
 95%|█████████▌| 1888/1987 [00:01<00:00, 997.66it/s]
+[2026-05-30 00:17:14,755][lm_eval.evaluator][INFO] - Running loglikelihood requests
+

data/ulars/supplement/standard_eval/logs/saisasanky__zephyr_rmu__wmdp_cyber.log

@@ -0,0 +1,22 @@
+repo_id: saisasanky/zephyr_rmu
+split: cyber
+snapshot: /zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/saisasanky__zephyr_rmu/models--saisasanky--zephyr_rmu/snapshots/cd0a255971286a18ffdedacb52fc8fca41398b3c
+cmd: /zfspool/zangyihe/conda_envs/corgo/bin/python src/eval.py experiment=eval/wmdp/default model=zephyr-7b-beta model.model_args.pretrained_model_name_or_path=/zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/saisasanky__zephyr_rmu/models--saisasanky--zephyr_rmu/snapshots/cd0a255971286a18ffdedacb52fc8fca41398b3c model.tokenizer_args.pretrained_model_name_or_path=/zfspool/zangyihe/ulars/models/zephyr-7b-beta model.model_args.attn_implementation=sdpa model.model_args.torch_dtype=bfloat16 data_split=cyber task_name=hf_wmdp_cyber_saisasanky__zephyr_rmu paths.output_dir=/zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/saisasanky__zephyr_rmu__wmdp_cyber eval.lm_eval.overwrite=true eval.lm_eval.tasks=[wmdp_cyber] eval.lm_eval.simple_evaluate_args.batch_size=16
+
+/zfspool/zangyihe/conda_envs/corgo/lib/python3.11/site-packages/transformers/utils/hub.py:105: FutureWarning: Using `TRANSFORMERS_CACHE` is deprecated and will be removed in v5 of Transformers. Use `HF_HOME` instead.
+  warnings.warn(
+[2026-05-30 00:17:38,113] [INFO] [real_accelerator.py:219:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+
+[2026-05-30 00:17:48,015][lm_eval.models.huggingface][WARNING] - `pretrained` model kwarg is not of type `str`. Many other model arguments may be ignored. Please do not launch via accelerate or use `parallelize=True` if passing an existing model this way.
+[2026-05-30 00:17:48,027][lm_eval.models.huggingface][WARNING] - Passed an already-initialized model through `pretrained`, assuming single-process call to evaluate() or custom distributed integration
+[2026-05-30 00:17:48,029][evaluator][INFO] - ***** Running LMEval evaluation suite *****
+[2026-05-30 00:17:48,029][evaluator][INFO] - Fine-grained evaluations will be saved to: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/saisasanky__zephyr_rmu__wmdp_cyber/LMEval_EVAL.json
+[2026-05-30 00:17:48,029][evaluator][INFO] - Aggregated evaluations will be summarised in: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/saisasanky__zephyr_rmu__wmdp_cyber/LMEval_SUMMARY.json
+[2026-05-30 00:17:48,031][lm_eval.evaluator][INFO] - Setting random seed to 0 | Setting numpy seed to 1234 | Setting torch manual seed to 1234 | Setting fewshot manual seed to 1234
+[2026-05-30 00:17:48,031][lm_eval.evaluator][INFO] - Using pre-initialized model
+[2026-05-30 00:17:52,084][lm_eval.tasks][INFO] - Selected tasks:
+[2026-05-30 00:17:52,085][lm_eval.tasks][INFO] - Task: wmdp_cyber (wmdp/wmdp_cyber.yaml)
+[2026-05-30 00:17:52,085][lm_eval.api.task][INFO] - Building contexts for wmdp_cyber on rank 0...
+
  0%|          | 0/1987 [00:00<?, ?it/s]
  4%|▍         | 80/1987 [00:00<00:02, 793.41it/s]
  9%|▉         | 177/1987 [00:00<00:02, 896.86it/s]
 14%|█▍        | 274/1987 [00:00<00:01, 928.18it/s]
 19%|█▊        | 372/1987 [00:00<00:01, 947.73it/s]
 24%|██▎       | 471/1987 [00:00<00:01, 961.64it/s]
 29%|██▊       | 568/1987 [00:00<00:01, 961.65it/s]
 34%|███▎      | 667/1987 [00:00<00:01, 970.12it/s]
 39%|███▊      | 765/1987 [00:00<00:01, 967.41it/s]
 43%|████▎     | 864/1987 [00:00<00:01, 973.98it/s]
 49%|████▊     | 964/1987 [00:01<00:01, 980.25it/s]
 54%|█████▎    | 1064/1987 [00:01<00:00, 984.09it/s]
 59%|█████▊    | 1164/1987 [00:01<00:00, 986.12it/s]
 64%|██████▎   | 1263/1987 [00:01<00:00, 980.00it/s]
 69%|██████▊   | 1362/1987 [00:01<00:00, 982.07it/s]
 74%|███████▎  | 1462/1987 [00:01<00:00, 985.56it/s]
 79%|███████▊  | 1562/1987 [00:01<00:00, 987.69it/s]
 84%|████████▎ | 1662/1987 [00:01<00:00, 989.34it/s]
 89%|████████▊ | 1761/1987 [00:01<00:00, 978.50it/s]
 94%|█████████▎| 1860/1987 [00:01<00:00, 981.17it/s]
 99%|█████████▊| 1960/1987 [00:02<00:00, 985.67it/s]
+[2026-05-30 00:17:54,179][lm_eval.evaluator][INFO] - Running loglikelihood requests
+

data/ulars/supplement/standard_eval/logs/shengyuanhu__benchmark_wmdp_ga_ckpt_140__wmdp_cyber.log

@@ -0,0 +1,22 @@
+repo_id: shengyuanhu/benchmark_wmdp_ga_ckpt_140
+split: cyber
+snapshot: /zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/shengyuanhu__benchmark_wmdp_ga_ckpt_140/models--shengyuanhu--benchmark_wmdp_ga_ckpt_140/snapshots/83097818b82ed600184309eb193bdb17de44a930
+cmd: /zfspool/zangyihe/conda_envs/corgo/bin/python src/eval.py experiment=eval/wmdp/default model=zephyr-7b-beta model.model_args.pretrained_model_name_or_path=/zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/shengyuanhu__benchmark_wmdp_ga_ckpt_140/models--shengyuanhu--benchmark_wmdp_ga_ckpt_140/snapshots/83097818b82ed600184309eb193bdb17de44a930 model.tokenizer_args.pretrained_model_name_or_path=/zfspool/zangyihe/ulars/models/zephyr-7b-beta model.model_args.attn_implementation=sdpa model.model_args.torch_dtype=bfloat16 data_split=cyber task_name=hf_wmdp_cyber_shengyuanhu__benchmark_wmdp_ga_ckpt_140 paths.output_dir=/zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/shengyuanhu__benchmark_wmdp_ga_ckpt_140__wmdp_cyber eval.lm_eval.overwrite=true eval.lm_eval.tasks=[wmdp_cyber] eval.lm_eval.simple_evaluate_args.batch_size=16
+
+/zfspool/zangyihe/conda_envs/corgo/lib/python3.11/site-packages/transformers/utils/hub.py:105: FutureWarning: Using `TRANSFORMERS_CACHE` is deprecated and will be removed in v5 of Transformers. Use `HF_HOME` instead.
+  warnings.warn(
+[2026-05-30 00:19:17,488] [INFO] [real_accelerator.py:219:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+
+[2026-05-30 00:19:32,861][lm_eval.models.huggingface][WARNING] - `pretrained` model kwarg is not of type `str`. Many other model arguments may be ignored. Please do not launch via accelerate or use `parallelize=True` if passing an existing model this way.
+[2026-05-30 00:19:32,873][lm_eval.models.huggingface][WARNING] - Passed an already-initialized model through `pretrained`, assuming single-process call to evaluate() or custom distributed integration
+[2026-05-30 00:19:32,874][evaluator][INFO] - ***** Running LMEval evaluation suite *****
+[2026-05-30 00:19:32,875][evaluator][INFO] - Fine-grained evaluations will be saved to: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/shengyuanhu__benchmark_wmdp_ga_ckpt_140__wmdp_cyber/LMEval_EVAL.json
+[2026-05-30 00:19:32,875][evaluator][INFO] - Aggregated evaluations will be summarised in: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/shengyuanhu__benchmark_wmdp_ga_ckpt_140__wmdp_cyber/LMEval_SUMMARY.json
+[2026-05-30 00:19:32,876][lm_eval.evaluator][INFO] - Setting random seed to 0 | Setting numpy seed to 1234 | Setting torch manual seed to 1234 | Setting fewshot manual seed to 1234
+[2026-05-30 00:19:32,876][lm_eval.evaluator][INFO] - Using pre-initialized model
+[2026-05-30 00:19:36,559][lm_eval.tasks][INFO] - Selected tasks:
+[2026-05-30 00:19:36,559][lm_eval.tasks][INFO] - Task: wmdp_cyber (wmdp/wmdp_cyber.yaml)
+[2026-05-30 00:19:36,559][lm_eval.api.task][INFO] - Building contexts for wmdp_cyber on rank 0...
+
  0%|          | 0/1987 [00:00<?, ?it/s]
  5%|▍         | 96/1987 [00:00<00:01, 958.84it/s]
 10%|▉         | 192/1987 [00:00<00:01, 939.78it/s]
 15%|█▍        | 291/1987 [00:00<00:01, 960.24it/s]
 20%|█▉        | 391/1987 [00:00<00:01, 972.74it/s]
 25%|██▍       | 490/1987 [00:00<00:01, 978.33it/s]
 30%|██▉       | 590/1987 [00:00<00:01, 984.29it/s]
 35%|███▍      | 690/1987 [00:00<00:01, 988.25it/s]
 40%|███▉      | 789/1987 [00:00<00:01, 975.28it/s]
 45%|████▍     | 889/1987 [00:00<00:01, 980.23it/s]
 50%|████▉     | 989/1987 [00:01<00:01, 984.80it/s]
 55%|█████▍    | 1089/1987 [00:01<00:00, 988.05it/s]
 60%|█████▉    | 1189/1987 [00:01<00:00, 991.27it/s]
 65%|██████▍   | 1289/1987 [00:01<00:00, 991.02it/s]
 70%|██████▉   | 1389/1987 [00:01<00:00, 990.93it/s]
 75%|███████▍  | 1489/1987 [00:01<00:00, 991.10it/s]
 80%|███████▉  | 1589/1987 [00:01<00:00, 992.73it/s]
 85%|████████▌ | 1690/1987 [00:01<00:00, 995.30it/s]
 90%|█████████ | 1790/1987 [00:01<00:00, 987.55it/s]
 95%|█████████▌| 1890/1987 [00:01<00:00, 990.82it/s]
+[2026-05-30 00:19:38,628][lm_eval.evaluator][INFO] - Running loglikelihood requests
+

data/ulars/supplement/standard_eval/logs/shengyuanhu__benchmark_wmdp_kl_ckpt_140__wmdp_bio.log

@@ -0,0 +1,22 @@
+repo_id: shengyuanhu/benchmark_wmdp_kl_ckpt_140
+split: bio
+snapshot: /zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/shengyuanhu__benchmark_wmdp_kl_ckpt_140/models--shengyuanhu--benchmark_wmdp_kl_ckpt_140/snapshots/97fc3794ea7b3df8f1cfce236bbacd36ad367503
+cmd: /zfspool/zangyihe/conda_envs/corgo/bin/python src/eval.py experiment=eval/wmdp/default model=zephyr-7b-beta model.model_args.pretrained_model_name_or_path=/zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/shengyuanhu__benchmark_wmdp_kl_ckpt_140/models--shengyuanhu--benchmark_wmdp_kl_ckpt_140/snapshots/97fc3794ea7b3df8f1cfce236bbacd36ad367503 model.tokenizer_args.pretrained_model_name_or_path=/zfspool/zangyihe/ulars/models/zephyr-7b-beta model.model_args.attn_implementation=sdpa model.model_args.torch_dtype=bfloat16 data_split=bio task_name=hf_wmdp_bio_shengyuanhu__benchmark_wmdp_kl_ckpt_140 paths.output_dir=/zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/shengyuanhu__benchmark_wmdp_kl_ckpt_140__wmdp_bio eval.lm_eval.overwrite=true eval.lm_eval.tasks=[wmdp_bio] eval.lm_eval.simple_evaluate_args.batch_size=16
+
+/zfspool/zangyihe/conda_envs/corgo/lib/python3.11/site-packages/transformers/utils/hub.py:105: FutureWarning: Using `TRANSFORMERS_CACHE` is deprecated and will be removed in v5 of Transformers. Use `HF_HOME` instead.
+  warnings.warn(
+[2026-05-29 22:09:45,336] [INFO] [real_accelerator.py:219:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+
+[2026-05-29 22:10:10,812][lm_eval.models.huggingface][WARNING] - `pretrained` model kwarg is not of type `str`. Many other model arguments may be ignored. Please do not launch via accelerate or use `parallelize=True` if passing an existing model this way.
+[2026-05-29 22:10:10,825][lm_eval.models.huggingface][WARNING] - Passed an already-initialized model through `pretrained`, assuming single-process call to evaluate() or custom distributed integration
+[2026-05-29 22:10:10,826][evaluator][INFO] - ***** Running LMEval evaluation suite *****
+[2026-05-29 22:10:10,826][evaluator][INFO] - Fine-grained evaluations will be saved to: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/shengyuanhu__benchmark_wmdp_kl_ckpt_140__wmdp_bio/LMEval_EVAL.json
+[2026-05-29 22:10:10,826][evaluator][INFO] - Aggregated evaluations will be summarised in: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/shengyuanhu__benchmark_wmdp_kl_ckpt_140__wmdp_bio/LMEval_SUMMARY.json
+[2026-05-29 22:10:10,827][lm_eval.evaluator][INFO] - Setting random seed to 0 | Setting numpy seed to 1234 | Setting torch manual seed to 1234 | Setting fewshot manual seed to 1234
+[2026-05-29 22:10:10,827][lm_eval.evaluator][INFO] - Using pre-initialized model
+[2026-05-29 22:10:20,988][lm_eval.tasks][INFO] - Selected tasks:
+[2026-05-29 22:10:20,988][lm_eval.tasks][INFO] - Task: wmdp_bio (wmdp/wmdp_bio.yaml)
+[2026-05-29 22:10:20,989][lm_eval.api.task][INFO] - Building contexts for wmdp_bio on rank 0...
+
  0%|          | 0/1273 [00:00<?, ?it/s]
  8%|▊         | 97/1273 [00:00<00:01, 966.42it/s]
 15%|█▌        | 196/1273 [00:00<00:01, 978.08it/s]
 23%|██▎       | 295/1273 [00:00<00:00, 979.99it/s]
 31%|███       | 394/1273 [00:00<00:00, 980.41it/s]
 39%|███▊      | 493/1273 [00:00<00:00, 982.40it/s]
 47%|████▋     | 592/1273 [00:00<00:00, 984.57it/s]
 54%|█████▍    | 692/1273 [00:00<00:00, 988.44it/s]
 62%|██████▏   | 792/1273 [00:00<00:00, 989.40it/s]
 70%|██████▉   | 891/1273 [00:00<00:00, 987.80it/s]
 78%|███████▊  | 990/1273 [00:01<00:00, 987.31it/s]
 86%|████████▌ | 1090/1273 [00:01<00:00, 990.63it/s]
 93%|█████████▎| 1190/1273 [00:01<00:00, 991.51it/s]
+[2026-05-29 22:10:22,309][lm_eval.evaluator][INFO] - Running loglikelihood requests
+

data/ulars/supplement/standard_eval/logs/shengyuanhu__benchmark_wmdp_kl_ckpt_140__wmdp_cyber.log

@@ -0,0 +1,26 @@
+repo_id: shengyuanhu/benchmark_wmdp_kl_ckpt_140
+split: cyber
+snapshot: /zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/shengyuanhu__benchmark_wmdp_kl_ckpt_140/models--shengyuanhu--benchmark_wmdp_kl_ckpt_140/snapshots/97fc3794ea7b3df8f1cfce236bbacd36ad367503
+cmd: /zfspool/zangyihe/conda_envs/corgo/bin/python src/eval.py experiment=eval/wmdp/default model=zephyr-7b-beta model.model_args.pretrained_model_name_or_path=/zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/shengyuanhu__benchmark_wmdp_kl_ckpt_140/models--shengyuanhu--benchmark_wmdp_kl_ckpt_140/snapshots/97fc3794ea7b3df8f1cfce236bbacd36ad367503 model.tokenizer_args.pretrained_model_name_or_path=/zfspool/zangyihe/ulars/models/zephyr-7b-beta model.model_args.attn_implementation=sdpa model.model_args.torch_dtype=bfloat16 data_split=cyber task_name=hf_wmdp_cyber_shengyuanhu__benchmark_wmdp_kl_ckpt_140 paths.output_dir=/zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/shengyuanhu__benchmark_wmdp_kl_ckpt_140__wmdp_cyber eval.lm_eval.overwrite=true eval.lm_eval.tasks=[wmdp_cyber] eval.lm_eval.simple_evaluate_args.batch_size=16
+
+/zfspool/zangyihe/conda_envs/corgo/lib/python3.11/site-packages/transformers/utils/hub.py:105: FutureWarning: Using `TRANSFORMERS_CACHE` is deprecated and will be removed in v5 of Transformers. Use `HF_HOME` instead.
+  warnings.warn(
+[2026-05-30 00:19:56,694] [INFO] [real_accelerator.py:219:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+
+[2026-05-30 00:20:54,890][lm_eval.models.huggingface][WARNING] - `pretrained` model kwarg is not of type `str`. Many other model arguments may be ignored. Please do not launch via accelerate or use `parallelize=True` if passing an existing model this way.
+[2026-05-30 00:20:54,902][lm_eval.models.huggingface][WARNING] - Passed an already-initialized model through `pretrained`, assuming single-process call to evaluate() or custom distributed integration
+[2026-05-30 00:20:54,903][evaluator][INFO] - ***** Running LMEval evaluation suite *****
+[2026-05-30 00:20:54,903][evaluator][INFO] - Fine-grained evaluations will be saved to: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/shengyuanhu__benchmark_wmdp_kl_ckpt_140__wmdp_cyber/LMEval_EVAL.json
+[2026-05-30 00:20:54,904][evaluator][INFO] - Aggregated evaluations will be summarised in: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/shengyuanhu__benchmark_wmdp_kl_ckpt_140__wmdp_cyber/LMEval_SUMMARY.json
+[2026-05-30 00:20:54,905][lm_eval.evaluator][INFO] - Setting random seed to 0 | Setting numpy seed to 1234 | Setting torch manual seed to 1234 | Setting fewshot manual seed to 1234
+[2026-05-30 00:20:54,905][lm_eval.evaluator][INFO] - Using pre-initialized model
+Using the latest cached version of the dataset since cais/wmdp couldn't be found on the Hugging Face Hub
+[2026-05-30 00:21:13,436][datasets.load][WARNING] - Using the latest cached version of the dataset since cais/wmdp couldn't be found on the Hugging Face Hub
+Found the latest cached dataset configuration 'wmdp-cyber' at /zfspool/zangyihe/ulars/hf-wmdp-eval/datasets-cache/cais___wmdp/wmdp-cyber/0.0.0/7125571f22f032c56415e7980f48d877dd830ff8 (last modified on Fri May 29 23:25:00 2026).
+[2026-05-30 00:21:13,437][datasets.packaged_modules.cache.cache][WARNING] - Found the latest cached dataset configuration 'wmdp-cyber' at /zfspool/zangyihe/ulars/hf-wmdp-eval/datasets-cache/cais___wmdp/wmdp-cyber/0.0.0/7125571f22f032c56415e7980f48d877dd830ff8 (last modified on Fri May 29 23:25:00 2026).
+[2026-05-30 00:21:13,499][lm_eval.tasks][INFO] - Selected tasks:
+[2026-05-30 00:21:13,499][lm_eval.tasks][INFO] - Task: wmdp_cyber (wmdp/wmdp_cyber.yaml)
+[2026-05-30 00:21:13,499][lm_eval.api.task][INFO] - Building contexts for wmdp_cyber on rank 0...
+
  0%|          | 0/1987 [00:00<?, ?it/s]
  5%|▍         | 97/1987 [00:00<00:01, 961.23it/s]
 10%|▉         | 195/1987 [00:00<00:01, 972.13it/s]
 15%|█▍        | 294/1987 [00:00<00:01, 976.69it/s]
 20%|█▉        | 392/1987 [00:00<00:01, 968.04it/s]
 25%|██▍       | 491/1987 [00:00<00:01, 973.40it/s]
 30%|██▉       | 590/1987 [00:00<00:01, 976.17it/s]
 35%|███▍      | 689/1987 [00:00<00:01, 979.77it/s]
 40%|███▉      | 789/1987 [00:00<00:01, 983.66it/s]
 45%|████▍     | 888/1987 [00:00<00:01, 981.29it/s]
 50%|████▉     | 988/1987 [00:01<00:01, 984.73it/s]
 55%|█████▍    | 1088/1987 [00:01<00:00, 987.41it/s]
 60%|█████▉    | 1187/1987 [00:01<00:00, 987.45it/s]
 65%|██████▍   | 1286/1987 [00:01<00:00, 988.06it/s]
 70%|██████▉   | 1385/1987 [00:01<00:00, 988.53it/s]
 75%|███████▍  | 1485/1987 [00:01<00:00, 990.03it/s]
 80%|███████▉  | 1585/1987 [00:01<00:00, 991.20it/s]
 85%|████████▍ | 1685/1987 [00:01<00:00, 988.79it/s]
 90%|████████▉ | 1785/1987 [00:01<00:00, 990.91it/s]
 95%|█████████▍| 1885/1987 [00:01<00:00, 991.63it/s]
+[2026-05-30 00:21:15,565][lm_eval.evaluator][INFO] - Running loglikelihood requests
+

data/ulars/supplement/standard_eval/logs/shengyuanhu__benchmark_wmdp_npo_kl_ckpt_75__wmdp_bio.log

@@ -0,0 +1,22 @@
+repo_id: shengyuanhu/benchmark_wmdp_npo_kl_ckpt_75
+split: bio
+snapshot: /zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/shengyuanhu__benchmark_wmdp_npo_kl_ckpt_75/models--shengyuanhu--benchmark_wmdp_npo_kl_ckpt_75/snapshots/1221875c099dcd1b32ba72237136428f9927aba0
+cmd: /zfspool/zangyihe/conda_envs/corgo/bin/python src/eval.py experiment=eval/wmdp/default model=zephyr-7b-beta model.model_args.pretrained_model_name_or_path=/zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/shengyuanhu__benchmark_wmdp_npo_kl_ckpt_75/models--shengyuanhu--benchmark_wmdp_npo_kl_ckpt_75/snapshots/1221875c099dcd1b32ba72237136428f9927aba0 model.tokenizer_args.pretrained_model_name_or_path=/zfspool/zangyihe/ulars/models/zephyr-7b-beta model.model_args.attn_implementation=sdpa model.model_args.torch_dtype=bfloat16 data_split=bio task_name=hf_wmdp_bio_shengyuanhu__benchmark_wmdp_npo_kl_ckpt_75 paths.output_dir=/zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/shengyuanhu__benchmark_wmdp_npo_kl_ckpt_75__wmdp_bio eval.lm_eval.overwrite=true eval.lm_eval.tasks=[wmdp_bio] eval.lm_eval.simple_evaluate_args.batch_size=16
+
+/zfspool/zangyihe/conda_envs/corgo/lib/python3.11/site-packages/transformers/utils/hub.py:105: FutureWarning: Using `TRANSFORMERS_CACHE` is deprecated and will be removed in v5 of Transformers. Use `HF_HOME` instead.
+  warnings.warn(
+[2026-05-29 22:32:03,120] [INFO] [real_accelerator.py:219:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+
+[2026-05-29 22:32:33,195][lm_eval.models.huggingface][WARNING] - `pretrained` model kwarg is not of type `str`. Many other model arguments may be ignored. Please do not launch via accelerate or use `parallelize=True` if passing an existing model this way.
+[2026-05-29 22:32:33,207][lm_eval.models.huggingface][WARNING] - Passed an already-initialized model through `pretrained`, assuming single-process call to evaluate() or custom distributed integration
+[2026-05-29 22:32:33,208][evaluator][INFO] - ***** Running LMEval evaluation suite *****
+[2026-05-29 22:32:33,208][evaluator][INFO] - Fine-grained evaluations will be saved to: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/shengyuanhu__benchmark_wmdp_npo_kl_ckpt_75__wmdp_bio/LMEval_EVAL.json
+[2026-05-29 22:32:33,208][evaluator][INFO] - Aggregated evaluations will be summarised in: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/shengyuanhu__benchmark_wmdp_npo_kl_ckpt_75__wmdp_bio/LMEval_SUMMARY.json
+[2026-05-29 22:32:33,209][lm_eval.evaluator][INFO] - Setting random seed to 0 | Setting numpy seed to 1234 | Setting torch manual seed to 1234 | Setting fewshot manual seed to 1234
+[2026-05-29 22:32:33,210][lm_eval.evaluator][INFO] - Using pre-initialized model
+[2026-05-29 22:32:42,473][lm_eval.tasks][INFO] - Selected tasks:
+[2026-05-29 22:32:42,473][lm_eval.tasks][INFO] - Task: wmdp_bio (wmdp/wmdp_bio.yaml)
+[2026-05-29 22:32:42,474][lm_eval.api.task][INFO] - Building contexts for wmdp_bio on rank 0...
+
  0%|          | 0/1273 [00:00<?, ?it/s]
  8%|▊         | 96/1273 [00:00<00:01, 959.51it/s]
 15%|█▌        | 195/1273 [00:00<00:01, 975.60it/s]
 23%|██▎       | 293/1273 [00:00<00:01, 974.90it/s]
 31%|███       | 392/1273 [00:00<00:00, 978.37it/s]
 38%|███▊      | 490/1273 [00:00<00:00, 977.89it/s]
 46%|████▋     | 589/1273 [00:00<00:00, 980.62it/s]
 54%|█████▍    | 689/1273 [00:00<00:00, 984.21it/s]
 62%|██████▏   | 789/1273 [00:00<00:00, 987.32it/s]
 70%|██████▉   | 889/1273 [00:00<00:00, 989.26it/s]
 78%|███████▊  | 989/1273 [00:01<00:00, 990.43it/s]
 86%|████████▌ | 1089/1273 [00:01<00:00, 992.31it/s]
 93%|█████████▎| 1189/1273 [00:01<00:00, 994.21it/s]
+[2026-05-29 22:32:43,795][lm_eval.evaluator][INFO] - Running loglikelihood requests
+

data/ulars/supplement/standard_eval/logs/shengyuanhu__benchmark_wmdp_scrub_ckpt_140__wmdp_bio.log

@@ -0,0 +1,22 @@
+repo_id: shengyuanhu/benchmark_wmdp_scrub_ckpt_140
+split: bio
+snapshot: /zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/shengyuanhu__benchmark_wmdp_scrub_ckpt_140/models--shengyuanhu--benchmark_wmdp_scrub_ckpt_140/snapshots/19fbf6f69bac5c7a85fb687dadaf16af6dc908e2
+cmd: /zfspool/zangyihe/conda_envs/corgo/bin/python src/eval.py experiment=eval/wmdp/default model=zephyr-7b-beta model.model_args.pretrained_model_name_or_path=/zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/shengyuanhu__benchmark_wmdp_scrub_ckpt_140/models--shengyuanhu--benchmark_wmdp_scrub_ckpt_140/snapshots/19fbf6f69bac5c7a85fb687dadaf16af6dc908e2 model.tokenizer_args.pretrained_model_name_or_path=/zfspool/zangyihe/ulars/models/zephyr-7b-beta model.model_args.attn_implementation=sdpa model.model_args.torch_dtype=bfloat16 data_split=bio task_name=hf_wmdp_bio_shengyuanhu__benchmark_wmdp_scrub_ckpt_140 paths.output_dir=/zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/shengyuanhu__benchmark_wmdp_scrub_ckpt_140__wmdp_bio eval.lm_eval.overwrite=true eval.lm_eval.tasks=[wmdp_bio] eval.lm_eval.simple_evaluate_args.batch_size=16
+
+/zfspool/zangyihe/conda_envs/corgo/lib/python3.11/site-packages/transformers/utils/hub.py:105: FutureWarning: Using `TRANSFORMERS_CACHE` is deprecated and will be removed in v5 of Transformers. Use `HF_HOME` instead.
+  warnings.warn(
+[2026-05-29 22:15:11,067] [INFO] [real_accelerator.py:219:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+
+[2026-05-29 22:15:36,026][lm_eval.models.huggingface][WARNING] - `pretrained` model kwarg is not of type `str`. Many other model arguments may be ignored. Please do not launch via accelerate or use `parallelize=True` if passing an existing model this way.
+[2026-05-29 22:15:36,039][lm_eval.models.huggingface][WARNING] - Passed an already-initialized model through `pretrained`, assuming single-process call to evaluate() or custom distributed integration
+[2026-05-29 22:15:36,040][evaluator][INFO] - ***** Running LMEval evaluation suite *****
+[2026-05-29 22:15:36,040][evaluator][INFO] - Fine-grained evaluations will be saved to: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/shengyuanhu__benchmark_wmdp_scrub_ckpt_140__wmdp_bio/LMEval_EVAL.json
+[2026-05-29 22:15:36,040][evaluator][INFO] - Aggregated evaluations will be summarised in: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/shengyuanhu__benchmark_wmdp_scrub_ckpt_140__wmdp_bio/LMEval_SUMMARY.json
+[2026-05-29 22:15:36,042][lm_eval.evaluator][INFO] - Setting random seed to 0 | Setting numpy seed to 1234 | Setting torch manual seed to 1234 | Setting fewshot manual seed to 1234
+[2026-05-29 22:15:36,042][lm_eval.evaluator][INFO] - Using pre-initialized model
+[2026-05-29 22:15:45,923][lm_eval.tasks][INFO] - Selected tasks:
+[2026-05-29 22:15:45,924][lm_eval.tasks][INFO] - Task: wmdp_bio (wmdp/wmdp_bio.yaml)
+[2026-05-29 22:15:45,924][lm_eval.api.task][INFO] - Building contexts for wmdp_bio on rank 0...
+
  0%|          | 0/1273 [00:00<?, ?it/s]
  8%|▊         | 97/1273 [00:00<00:01, 963.19it/s]
 15%|█▌        | 196/1273 [00:00<00:01, 978.33it/s]
 23%|██▎       | 296/1273 [00:00<00:00, 985.12it/s]
 31%|███       | 396/1273 [00:00<00:00, 990.14it/s]
 39%|███▉      | 496/1273 [00:00<00:00, 992.88it/s]
 47%|████▋     | 596/1273 [00:00<00:00, 995.13it/s]
 55%|█████▍    | 696/1273 [00:00<00:00, 995.55it/s]
 63%|██████▎   | 796/1273 [00:00<00:00, 988.70it/s]
 70%|███████   | 896/1273 [00:00<00:00, 989.78it/s]
 78%|███████▊  | 996/1273 [00:01<00:00, 992.12it/s]
 86%|████████▌ | 1097/1273 [00:01<00:00, 994.91it/s]
 94%|█████████▍| 1197/1273 [00:01<00:00, 995.55it/s]
+[2026-05-29 22:15:47,238][lm_eval.evaluator][INFO] - Running loglikelihood requests
+

data/ulars/supplement/standard_eval/logs/shengyuanhu__benchmark_wmdp_scrub_ckpt_140__wmdp_cyber.log

@@ -0,0 +1,22 @@
+repo_id: shengyuanhu/benchmark_wmdp_scrub_ckpt_140
+split: cyber
+snapshot: /zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/shengyuanhu__benchmark_wmdp_scrub_ckpt_140/models--shengyuanhu--benchmark_wmdp_scrub_ckpt_140/snapshots/19fbf6f69bac5c7a85fb687dadaf16af6dc908e2
+cmd: /zfspool/zangyihe/conda_envs/corgo/bin/python src/eval.py experiment=eval/wmdp/default model=zephyr-7b-beta model.model_args.pretrained_model_name_or_path=/zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/shengyuanhu__benchmark_wmdp_scrub_ckpt_140/models--shengyuanhu--benchmark_wmdp_scrub_ckpt_140/snapshots/19fbf6f69bac5c7a85fb687dadaf16af6dc908e2 model.tokenizer_args.pretrained_model_name_or_path=/zfspool/zangyihe/ulars/models/zephyr-7b-beta model.model_args.attn_implementation=sdpa model.model_args.torch_dtype=bfloat16 data_split=cyber task_name=hf_wmdp_cyber_shengyuanhu__benchmark_wmdp_scrub_ckpt_140 paths.output_dir=/zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/shengyuanhu__benchmark_wmdp_scrub_ckpt_140__wmdp_cyber eval.lm_eval.overwrite=true eval.lm_eval.tasks=[wmdp_cyber] eval.lm_eval.simple_evaluate_args.batch_size=16
+
+/zfspool/zangyihe/conda_envs/corgo/lib/python3.11/site-packages/transformers/utils/hub.py:105: FutureWarning: Using `TRANSFORMERS_CACHE` is deprecated and will be removed in v5 of Transformers. Use `HF_HOME` instead.
+  warnings.warn(
+[2026-05-30 00:23:17,007] [INFO] [real_accelerator.py:219:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+
+[2026-05-30 00:23:54,822][lm_eval.models.huggingface][WARNING] - `pretrained` model kwarg is not of type `str`. Many other model arguments may be ignored. Please do not launch via accelerate or use `parallelize=True` if passing an existing model this way.
+[2026-05-30 00:23:54,835][lm_eval.models.huggingface][WARNING] - Passed an already-initialized model through `pretrained`, assuming single-process call to evaluate() or custom distributed integration
+[2026-05-30 00:23:54,837][evaluator][INFO] - ***** Running LMEval evaluation suite *****
+[2026-05-30 00:23:54,837][evaluator][INFO] - Fine-grained evaluations will be saved to: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/shengyuanhu__benchmark_wmdp_scrub_ckpt_140__wmdp_cyber/LMEval_EVAL.json
+[2026-05-30 00:23:54,837][evaluator][INFO] - Aggregated evaluations will be summarised in: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/shengyuanhu__benchmark_wmdp_scrub_ckpt_140__wmdp_cyber/LMEval_SUMMARY.json
+[2026-05-30 00:23:54,838][lm_eval.evaluator][INFO] - Setting random seed to 0 | Setting numpy seed to 1234 | Setting torch manual seed to 1234 | Setting fewshot manual seed to 1234
+[2026-05-30 00:23:54,838][lm_eval.evaluator][INFO] - Using pre-initialized model
+[2026-05-30 00:24:01,941][lm_eval.tasks][INFO] - Selected tasks:
+[2026-05-30 00:24:01,942][lm_eval.tasks][INFO] - Task: wmdp_cyber (wmdp/wmdp_cyber.yaml)
+[2026-05-30 00:24:01,942][lm_eval.api.task][INFO] - Building contexts for wmdp_cyber on rank 0...
+
  0%|          | 0/1987 [00:00<?, ?it/s]
  5%|▍         | 95/1987 [00:00<00:02, 944.15it/s]
 10%|▉         | 192/1987 [00:00<00:01, 959.14it/s]
 15%|█▍        | 290/1987 [00:00<00:01, 964.83it/s]
 20%|█▉        | 389/1987 [00:00<00:01, 971.73it/s]
 25%|██▍       | 488/1987 [00:00<00:01, 975.66it/s]
 30%|██▉       | 587/1987 [00:00<00:01, 978.31it/s]
 35%|███▍      | 686/1987 [00:00<00:01, 981.76it/s]
 40%|███▉      | 785/1987 [00:00<00:01, 983.95it/s]
 44%|████▍     | 884/1987 [00:00<00:01, 984.08it/s]
 49%|████▉     | 983/1987 [00:01<00:01, 964.99it/s]
 54%|█████▍    | 1080/1987 [00:01<00:00, 963.98it/s]
 59%|█████▉    | 1179/1987 [00:01<00:00, 968.89it/s]
 64%|██████▍   | 1279/1987 [00:01<00:00, 975.78it/s]
 69%|██████▉   | 1379/1987 [00:01<00:00, 980.37it/s]
 74%|███████▍  | 1479/1987 [00:01<00:00, 983.44it/s]
 79%|███████▉  | 1578/1987 [00:01<00:00, 981.98it/s]
 84%|████████▍ | 1678/1987 [00:01<00:00, 985.38it/s]
 90%|████████▉ | 1779/1987 [00:01<00:00, 989.85it/s]
 95%|█████████▍| 1879/1987 [00:01<00:00, 991.16it/s]
+[2026-05-30 00:24:04,021][lm_eval.evaluator][INFO] - Running loglikelihood requests
+

data/ulars/supplement/standard_eval/logs/shengyuanhu__wmdp_unlearn_ga_ckpt_100_zephyr__wmdp_bio.log

@@ -0,0 +1,22 @@
+repo_id: shengyuanhu/wmdp_unlearn_ga_ckpt_100_zephyr
+split: bio
+snapshot: /zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/shengyuanhu__wmdp_unlearn_ga_ckpt_100_zephyr/models--shengyuanhu--wmdp_unlearn_ga_ckpt_100_zephyr/snapshots/39302ef8d1208623b66fc803bef8b47bbf7f86c7
+cmd: /zfspool/zangyihe/conda_envs/corgo/bin/python src/eval.py experiment=eval/wmdp/default model=zephyr-7b-beta model.model_args.pretrained_model_name_or_path=/zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/shengyuanhu__wmdp_unlearn_ga_ckpt_100_zephyr/models--shengyuanhu--wmdp_unlearn_ga_ckpt_100_zephyr/snapshots/39302ef8d1208623b66fc803bef8b47bbf7f86c7 model.tokenizer_args.pretrained_model_name_or_path=/zfspool/zangyihe/ulars/models/zephyr-7b-beta model.model_args.attn_implementation=sdpa model.model_args.torch_dtype=bfloat16 data_split=bio task_name=hf_wmdp_bio_shengyuanhu__wmdp_unlearn_ga_ckpt_100_zephyr paths.output_dir=/zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/shengyuanhu__wmdp_unlearn_ga_ckpt_100_zephyr__wmdp_bio eval.lm_eval.overwrite=true eval.lm_eval.tasks=[wmdp_bio] eval.lm_eval.simple_evaluate_args.batch_size=16
+
+/zfspool/zangyihe/conda_envs/corgo/lib/python3.11/site-packages/transformers/utils/hub.py:105: FutureWarning: Using `TRANSFORMERS_CACHE` is deprecated and will be removed in v5 of Transformers. Use `HF_HOME` instead.
+  warnings.warn(
+[2026-05-29 22:16:01,920] [INFO] [real_accelerator.py:219:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+
+[2026-05-29 22:16:33,681][lm_eval.models.huggingface][WARNING] - `pretrained` model kwarg is not of type `str`. Many other model arguments may be ignored. Please do not launch via accelerate or use `parallelize=True` if passing an existing model this way.
+[2026-05-29 22:16:33,694][lm_eval.models.huggingface][WARNING] - Passed an already-initialized model through `pretrained`, assuming single-process call to evaluate() or custom distributed integration
+[2026-05-29 22:16:33,695][evaluator][INFO] - ***** Running LMEval evaluation suite *****
+[2026-05-29 22:16:33,695][evaluator][INFO] - Fine-grained evaluations will be saved to: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/shengyuanhu__wmdp_unlearn_ga_ckpt_100_zephyr__wmdp_bio/LMEval_EVAL.json
+[2026-05-29 22:16:33,695][evaluator][INFO] - Aggregated evaluations will be summarised in: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/shengyuanhu__wmdp_unlearn_ga_ckpt_100_zephyr__wmdp_bio/LMEval_SUMMARY.json
+[2026-05-29 22:16:33,697][lm_eval.evaluator][INFO] - Setting random seed to 0 | Setting numpy seed to 1234 | Setting torch manual seed to 1234 | Setting fewshot manual seed to 1234
+[2026-05-29 22:16:33,697][lm_eval.evaluator][INFO] - Using pre-initialized model
+[2026-05-29 22:16:39,437][lm_eval.tasks][INFO] - Selected tasks:
+[2026-05-29 22:16:39,438][lm_eval.tasks][INFO] - Task: wmdp_bio (wmdp/wmdp_bio.yaml)
+[2026-05-29 22:16:39,438][lm_eval.api.task][INFO] - Building contexts for wmdp_bio on rank 0...
+
  0%|          | 0/1273 [00:00<?, ?it/s]
  7%|▋         | 95/1273 [00:00<00:01, 948.53it/s]
 15%|█▌        | 193/1273 [00:00<00:01, 964.02it/s]
 23%|██▎       | 290/1273 [00:00<00:01, 916.13it/s]
 30%|███       | 382/1273 [00:00<00:00, 908.78it/s]
 37%|███▋      | 474/1273 [00:00<00:00, 877.67it/s]
 44%|████▍     | 562/1273 [00:00<00:00, 865.98it/s]
 51%|█████     | 649/1273 [00:00<00:00, 843.41it/s]
 58%|█████▊    | 734/1273 [00:00<00:00, 816.20it/s]
 65%|██████▍   | 827/1273 [00:00<00:00, 848.01it/s]
 73%|███████▎  | 923/1273 [00:01<00:00, 881.05it/s]
 80%|███████▉  | 1015/1273 [00:01<00:00, 891.19it/s]
 87%|████████▋ | 1112/1273 [00:01<00:00, 912.95it/s]
 95%|█████████▍| 1205/1273 [00:01<00:00, 915.62it/s]
+[2026-05-29 22:16:40,894][lm_eval.evaluator][INFO] - Running loglikelihood requests
+

data/ulars/supplement/standard_eval/logs/shengyuanhu__wmdp_unlearn_ga_ckpt_100_zephyr__wmdp_cyber.log

@@ -0,0 +1,22 @@
+repo_id: shengyuanhu/wmdp_unlearn_ga_ckpt_100_zephyr
+split: cyber
+snapshot: /zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/shengyuanhu__wmdp_unlearn_ga_ckpt_100_zephyr/models--shengyuanhu--wmdp_unlearn_ga_ckpt_100_zephyr/snapshots/39302ef8d1208623b66fc803bef8b47bbf7f86c7
+cmd: /zfspool/zangyihe/conda_envs/corgo/bin/python src/eval.py experiment=eval/wmdp/default model=zephyr-7b-beta model.model_args.pretrained_model_name_or_path=/zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/shengyuanhu__wmdp_unlearn_ga_ckpt_100_zephyr/models--shengyuanhu--wmdp_unlearn_ga_ckpt_100_zephyr/snapshots/39302ef8d1208623b66fc803bef8b47bbf7f86c7 model.tokenizer_args.pretrained_model_name_or_path=/zfspool/zangyihe/ulars/models/zephyr-7b-beta model.model_args.attn_implementation=sdpa model.model_args.torch_dtype=bfloat16 data_split=cyber task_name=hf_wmdp_cyber_shengyuanhu__wmdp_unlearn_ga_ckpt_100_zephyr paths.output_dir=/zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/shengyuanhu__wmdp_unlearn_ga_ckpt_100_zephyr__wmdp_cyber eval.lm_eval.overwrite=true eval.lm_eval.tasks=[wmdp_cyber] eval.lm_eval.simple_evaluate_args.batch_size=16
+
+/zfspool/zangyihe/conda_envs/corgo/lib/python3.11/site-packages/transformers/utils/hub.py:105: FutureWarning: Using `TRANSFORMERS_CACHE` is deprecated and will be removed in v5 of Transformers. Use `HF_HOME` instead.
+  warnings.warn(
+[2026-05-30 00:49:11,797] [INFO] [real_accelerator.py:219:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+
+[2026-05-30 00:49:19,014][lm_eval.models.huggingface][WARNING] - `pretrained` model kwarg is not of type `str`. Many other model arguments may be ignored. Please do not launch via accelerate or use `parallelize=True` if passing an existing model this way.
+[2026-05-30 00:49:19,026][lm_eval.models.huggingface][WARNING] - Passed an already-initialized model through `pretrained`, assuming single-process call to evaluate() or custom distributed integration
+[2026-05-30 00:49:19,028][evaluator][INFO] - ***** Running LMEval evaluation suite *****
+[2026-05-30 00:49:19,028][evaluator][INFO] - Fine-grained evaluations will be saved to: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/shengyuanhu__wmdp_unlearn_ga_ckpt_100_zephyr__wmdp_cyber/LMEval_EVAL.json
+[2026-05-30 00:49:19,028][evaluator][INFO] - Aggregated evaluations will be summarised in: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/shengyuanhu__wmdp_unlearn_ga_ckpt_100_zephyr__wmdp_cyber/LMEval_SUMMARY.json
+[2026-05-30 00:49:19,029][lm_eval.evaluator][INFO] - Setting random seed to 0 | Setting numpy seed to 1234 | Setting torch manual seed to 1234 | Setting fewshot manual seed to 1234
+[2026-05-30 00:49:19,029][lm_eval.evaluator][INFO] - Using pre-initialized model
+[2026-05-30 00:49:24,142][lm_eval.tasks][INFO] - Selected tasks:
+[2026-05-30 00:49:24,142][lm_eval.tasks][INFO] - Task: wmdp_cyber (wmdp/wmdp_cyber.yaml)
+[2026-05-30 00:49:24,143][lm_eval.api.task][INFO] - Building contexts for wmdp_cyber on rank 0...
+
  0%|          | 0/1987 [00:00<?, ?it/s]
  5%|▍         | 94/1987 [00:00<00:02, 938.34it/s]
 10%|▉         | 193/1987 [00:00<00:01, 963.65it/s]
 15%|█▍        | 292/1987 [00:00<00:01, 974.77it/s]
 20%|█▉        | 391/1987 [00:00<00:01, 979.97it/s]
 25%|██▍       | 491/1987 [00:00<00:01, 986.04it/s]
 30%|██▉       | 590/1987 [00:00<00:01, 980.12it/s]
 35%|███▍      | 690/1987 [00:00<00:01, 984.36it/s]
 40%|███▉      | 789/1987 [00:00<00:01, 985.47it/s]
 45%|████▍     | 889/1987 [00:00<00:01, 988.22it/s]
 50%|████▉     | 988/1987 [00:01<00:01, 956.10it/s]
 55%|█████▍    | 1088/1987 [00:01<00:00, 966.39it/s]
 60%|█████▉    | 1188/1987 [00:01<00:00, 975.57it/s]
 65%|██████▍   | 1289/1987 [00:01<00:00, 983.72it/s]
 70%|██████▉   | 1390/1987 [00:01<00:00, 989.10it/s]
 75%|███████▌  | 1491/1987 [00:01<00:00, 993.65it/s]
 80%|████████  | 1591/1987 [00:01<00:00, 970.05it/s]
 85%|████████▌ | 1689/1987 [00:01<00:00, 972.26it/s]
 90%|█████████ | 1790/1987 [00:01<00:00, 981.59it/s]
 95%|█████████▌| 1890/1987 [00:01<00:00, 986.38it/s]
+[2026-05-30 00:49:26,219][lm_eval.evaluator][INFO] - Running loglikelihood requests
+

data/ulars/supplement/standard_eval/logs/shengyuanhu__wmdp_unlearn_ga_ckpt_120_zephyr__wmdp_bio.log

@@ -0,0 +1,22 @@
+repo_id: shengyuanhu/wmdp_unlearn_ga_ckpt_120_zephyr
+split: bio
+snapshot: /zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/shengyuanhu__wmdp_unlearn_ga_ckpt_120_zephyr/models--shengyuanhu--wmdp_unlearn_ga_ckpt_120_zephyr/snapshots/1914642cdc7f50aa827164903bc67eadf4e948a3
+cmd: /zfspool/zangyihe/conda_envs/corgo/bin/python src/eval.py experiment=eval/wmdp/default model=zephyr-7b-beta model.model_args.pretrained_model_name_or_path=/zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/shengyuanhu__wmdp_unlearn_ga_ckpt_120_zephyr/models--shengyuanhu--wmdp_unlearn_ga_ckpt_120_zephyr/snapshots/1914642cdc7f50aa827164903bc67eadf4e948a3 model.tokenizer_args.pretrained_model_name_or_path=/zfspool/zangyihe/ulars/models/zephyr-7b-beta model.model_args.attn_implementation=sdpa model.model_args.torch_dtype=bfloat16 data_split=bio task_name=hf_wmdp_bio_shengyuanhu__wmdp_unlearn_ga_ckpt_120_zephyr paths.output_dir=/zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/shengyuanhu__wmdp_unlearn_ga_ckpt_120_zephyr__wmdp_bio eval.lm_eval.overwrite=true eval.lm_eval.tasks=[wmdp_bio] eval.lm_eval.simple_evaluate_args.batch_size=16
+
+/zfspool/zangyihe/conda_envs/corgo/lib/python3.11/site-packages/transformers/utils/hub.py:105: FutureWarning: Using `TRANSFORMERS_CACHE` is deprecated and will be removed in v5 of Transformers. Use `HF_HOME` instead.
+  warnings.warn(
+[2026-05-29 22:25:52,502] [INFO] [real_accelerator.py:219:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+
+[2026-05-29 22:26:20,531][lm_eval.models.huggingface][WARNING] - `pretrained` model kwarg is not of type `str`. Many other model arguments may be ignored. Please do not launch via accelerate or use `parallelize=True` if passing an existing model this way.
+[2026-05-29 22:26:20,543][lm_eval.models.huggingface][WARNING] - Passed an already-initialized model through `pretrained`, assuming single-process call to evaluate() or custom distributed integration
+[2026-05-29 22:26:20,544][evaluator][INFO] - ***** Running LMEval evaluation suite *****
+[2026-05-29 22:26:20,544][evaluator][INFO] - Fine-grained evaluations will be saved to: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/shengyuanhu__wmdp_unlearn_ga_ckpt_120_zephyr__wmdp_bio/LMEval_EVAL.json
+[2026-05-29 22:26:20,545][evaluator][INFO] - Aggregated evaluations will be summarised in: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/shengyuanhu__wmdp_unlearn_ga_ckpt_120_zephyr__wmdp_bio/LMEval_SUMMARY.json
+[2026-05-29 22:26:20,546][lm_eval.evaluator][INFO] - Setting random seed to 0 | Setting numpy seed to 1234 | Setting torch manual seed to 1234 | Setting fewshot manual seed to 1234
+[2026-05-29 22:26:20,546][lm_eval.evaluator][INFO] - Using pre-initialized model
+[2026-05-29 22:26:28,750][lm_eval.tasks][INFO] - Selected tasks:
+[2026-05-29 22:26:28,750][lm_eval.tasks][INFO] - Task: wmdp_bio (wmdp/wmdp_bio.yaml)
+[2026-05-29 22:26:28,750][lm_eval.api.task][INFO] - Building contexts for wmdp_bio on rank 0...
+
  0%|          | 0/1273 [00:00<?, ?it/s]
  8%|▊         | 96/1273 [00:00<00:01, 952.57it/s]
 15%|█▌        | 194/1273 [00:00<00:01, 968.44it/s]
 23%|██▎       | 291/1273 [00:00<00:01, 965.26it/s]
 31%|███       | 389/1273 [00:00<00:00, 969.88it/s]
 38%|███▊      | 488/1273 [00:00<00:00, 975.73it/s]
 46%|████▌     | 587/1273 [00:00<00:00, 979.59it/s]
 54%|█████▍    | 687/1273 [00:00<00:00, 983.22it/s]
 62%|██████▏   | 786/1273 [00:00<00:00, 984.76it/s]
 70%|██████▉   | 885/1273 [00:00<00:00, 985.18it/s]
 77%|███████▋  | 984/1273 [00:01<00:00, 985.48it/s]
 85%|████████▌ | 1083/1273 [00:01<00:00, 985.03it/s]
 93%|█████████▎| 1182/1273 [00:01<00:00, 986.11it/s]
+[2026-05-29 22:26:30,079][lm_eval.evaluator][INFO] - Running loglikelihood requests
+

data/ulars/supplement/standard_eval/logs/shengyuanhu__wmdp_unlearn_ga_ckpt_160_zephyr__wmdp_bio.log

@@ -0,0 +1,22 @@
+repo_id: shengyuanhu/wmdp_unlearn_ga_ckpt_160_zephyr
+split: bio
+snapshot: /zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/shengyuanhu__wmdp_unlearn_ga_ckpt_160_zephyr/models--shengyuanhu--wmdp_unlearn_ga_ckpt_160_zephyr/snapshots/1ee2304db8d55860529ac40105a6a8c8015b630b
+cmd: /zfspool/zangyihe/conda_envs/corgo/bin/python src/eval.py experiment=eval/wmdp/default model=zephyr-7b-beta model.model_args.pretrained_model_name_or_path=/zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/shengyuanhu__wmdp_unlearn_ga_ckpt_160_zephyr/models--shengyuanhu--wmdp_unlearn_ga_ckpt_160_zephyr/snapshots/1ee2304db8d55860529ac40105a6a8c8015b630b model.tokenizer_args.pretrained_model_name_or_path=/zfspool/zangyihe/ulars/models/zephyr-7b-beta model.model_args.attn_implementation=sdpa model.model_args.torch_dtype=bfloat16 data_split=bio task_name=hf_wmdp_bio_shengyuanhu__wmdp_unlearn_ga_ckpt_160_zephyr paths.output_dir=/zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/shengyuanhu__wmdp_unlearn_ga_ckpt_160_zephyr__wmdp_bio eval.lm_eval.overwrite=true eval.lm_eval.tasks=[wmdp_bio] eval.lm_eval.simple_evaluate_args.batch_size=16
+
+/zfspool/zangyihe/conda_envs/corgo/lib/python3.11/site-packages/transformers/utils/hub.py:105: FutureWarning: Using `TRANSFORMERS_CACHE` is deprecated and will be removed in v5 of Transformers. Use `HF_HOME` instead.
+  warnings.warn(
+[2026-05-29 22:25:24,859] [INFO] [real_accelerator.py:219:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+
+[2026-05-29 22:25:51,066][lm_eval.models.huggingface][WARNING] - `pretrained` model kwarg is not of type `str`. Many other model arguments may be ignored. Please do not launch via accelerate or use `parallelize=True` if passing an existing model this way.
+[2026-05-29 22:25:51,079][lm_eval.models.huggingface][WARNING] - Passed an already-initialized model through `pretrained`, assuming single-process call to evaluate() or custom distributed integration
+[2026-05-29 22:25:51,080][evaluator][INFO] - ***** Running LMEval evaluation suite *****
+[2026-05-29 22:25:51,080][evaluator][INFO] - Fine-grained evaluations will be saved to: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/shengyuanhu__wmdp_unlearn_ga_ckpt_160_zephyr__wmdp_bio/LMEval_EVAL.json
+[2026-05-29 22:25:51,080][evaluator][INFO] - Aggregated evaluations will be summarised in: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/shengyuanhu__wmdp_unlearn_ga_ckpt_160_zephyr__wmdp_bio/LMEval_SUMMARY.json
+[2026-05-29 22:25:51,081][lm_eval.evaluator][INFO] - Setting random seed to 0 | Setting numpy seed to 1234 | Setting torch manual seed to 1234 | Setting fewshot manual seed to 1234
+[2026-05-29 22:25:51,081][lm_eval.evaluator][INFO] - Using pre-initialized model
+[2026-05-29 22:25:58,116][lm_eval.tasks][INFO] - Selected tasks:
+[2026-05-29 22:25:58,116][lm_eval.tasks][INFO] - Task: wmdp_bio (wmdp/wmdp_bio.yaml)
+[2026-05-29 22:25:58,117][lm_eval.api.task][INFO] - Building contexts for wmdp_bio on rank 0...
+
  0%|          | 0/1273 [00:00<?, ?it/s]
  8%|▊         | 96/1273 [00:00<00:01, 951.13it/s]
 15%|█▌        | 193/1273 [00:00<00:01, 957.29it/s]
 23%|██▎       | 289/1273 [00:00<00:01, 957.60it/s]
 30%|███       | 387/1273 [00:00<00:00, 965.53it/s]
 38%|███▊      | 486/1273 [00:00<00:00, 971.21it/s]
 46%|████▌     | 585/1273 [00:00<00:00, 976.09it/s]
 54%|█████▎    | 684/1273 [00:00<00:00, 980.11it/s]
 62%|██████▏   | 783/1273 [00:00<00:00, 982.38it/s]
 69%|██████▉   | 882/1273 [00:00<00:00, 983.07it/s]
 77%|███████▋  | 981/1273 [00:01<00:00, 974.53it/s]
 85%|████████▍ | 1080/1273 [00:01<00:00, 977.60it/s]
 93%|█████████▎| 1179/1273 [00:01<00:00, 979.98it/s]
+[2026-05-29 22:25:59,452][lm_eval.evaluator][INFO] - Running loglikelihood requests
+

data/ulars/supplement/standard_eval/logs/shengyuanhu__wmdp_unlearn_gd_ckpt_100_zephyr__wmdp_cyber.log

@@ -0,0 +1,22 @@
+repo_id: shengyuanhu/wmdp_unlearn_gd_ckpt_100_zephyr
+split: cyber
+snapshot: /zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/shengyuanhu__wmdp_unlearn_gd_ckpt_100_zephyr/models--shengyuanhu--wmdp_unlearn_gd_ckpt_100_zephyr/snapshots/5a27460c91d3dbbbf903f8c2e65cd1fdd5c14a13
+cmd: /zfspool/zangyihe/conda_envs/corgo/bin/python src/eval.py experiment=eval/wmdp/default model=zephyr-7b-beta model.model_args.pretrained_model_name_or_path=/zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/shengyuanhu__wmdp_unlearn_gd_ckpt_100_zephyr/models--shengyuanhu--wmdp_unlearn_gd_ckpt_100_zephyr/snapshots/5a27460c91d3dbbbf903f8c2e65cd1fdd5c14a13 model.tokenizer_args.pretrained_model_name_or_path=/zfspool/zangyihe/ulars/models/zephyr-7b-beta model.model_args.attn_implementation=sdpa model.model_args.torch_dtype=bfloat16 data_split=cyber task_name=hf_wmdp_cyber_shengyuanhu__wmdp_unlearn_gd_ckpt_100_zephyr paths.output_dir=/zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/shengyuanhu__wmdp_unlearn_gd_ckpt_100_zephyr__wmdp_cyber eval.lm_eval.overwrite=true eval.lm_eval.tasks=[wmdp_cyber] eval.lm_eval.simple_evaluate_args.batch_size=16
+
+/zfspool/zangyihe/conda_envs/corgo/lib/python3.11/site-packages/transformers/utils/hub.py:105: FutureWarning: Using `TRANSFORMERS_CACHE` is deprecated and will be removed in v5 of Transformers. Use `HF_HOME` instead.
+  warnings.warn(
+[2026-05-30 00:26:27,071] [INFO] [real_accelerator.py:219:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+
+[2026-05-30 00:26:36,955][lm_eval.models.huggingface][WARNING] - `pretrained` model kwarg is not of type `str`. Many other model arguments may be ignored. Please do not launch via accelerate or use `parallelize=True` if passing an existing model this way.
+[2026-05-30 00:26:36,972][lm_eval.models.huggingface][WARNING] - Passed an already-initialized model through `pretrained`, assuming single-process call to evaluate() or custom distributed integration
+[2026-05-30 00:26:36,973][evaluator][INFO] - ***** Running LMEval evaluation suite *****
+[2026-05-30 00:26:36,973][evaluator][INFO] - Fine-grained evaluations will be saved to: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/shengyuanhu__wmdp_unlearn_gd_ckpt_100_zephyr__wmdp_cyber/LMEval_EVAL.json
+[2026-05-30 00:26:36,973][evaluator][INFO] - Aggregated evaluations will be summarised in: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/shengyuanhu__wmdp_unlearn_gd_ckpt_100_zephyr__wmdp_cyber/LMEval_SUMMARY.json
+[2026-05-30 00:26:36,974][lm_eval.evaluator][INFO] - Setting random seed to 0 | Setting numpy seed to 1234 | Setting torch manual seed to 1234 | Setting fewshot manual seed to 1234
+[2026-05-30 00:26:36,975][lm_eval.evaluator][INFO] - Using pre-initialized model
+[2026-05-30 00:26:42,747][lm_eval.tasks][INFO] - Selected tasks:
+[2026-05-30 00:26:42,747][lm_eval.tasks][INFO] - Task: wmdp_cyber (wmdp/wmdp_cyber.yaml)
+[2026-05-30 00:26:42,748][lm_eval.api.task][INFO] - Building contexts for wmdp_cyber on rank 0...
+
  0%|          | 0/1987 [00:00<?, ?it/s]
  5%|▍         | 96/1987 [00:00<00:01, 958.40it/s]
 10%|▉         | 195/1987 [00:00<00:01, 971.64it/s]
 15%|█▍        | 295/1987 [00:00<00:01, 980.80it/s]
 20%|█▉        | 394/1987 [00:00<00:01, 978.76it/s]
 25%|██▍       | 493/1987 [00:00<00:01, 982.12it/s]
 30%|██▉       | 593/1987 [00:00<00:01, 985.65it/s]
 35%|███▍      | 694/1987 [00:00<00:01, 990.46it/s]
 40%|███▉      | 794/1987 [00:00<00:01, 993.31it/s]
 45%|████▍     | 894/1987 [00:00<00:01, 995.07it/s]
 50%|█████     | 994/1987 [00:01<00:01, 895.85it/s]
 55%|█████▌    | 1094/1987 [00:01<00:00, 924.46it/s]
 60%|██████    | 1194/1987 [00:01<00:00, 944.53it/s]
 65%|██████▌   | 1294/1987 [00:01<00:00, 960.61it/s]
 70%|███████   | 1394/1987 [00:01<00:00, 970.83it/s]
 75%|███████▌  | 1493/1987 [00:01<00:00, 976.12it/s]
 80%|████████  | 1593/1987 [00:01<00:00, 982.16it/s]
 85%|████████▌ | 1693/1987 [00:01<00:00, 986.87it/s]
 90%|█████████ | 1794/1987 [00:01<00:00, 990.76it/s]
 95%|█████████▌| 1894/1987 [00:01<00:00, 992.42it/s]
+[2026-05-30 00:26:44,840][lm_eval.evaluator][INFO] - Running loglikelihood requests
+

data/ulars/supplement/standard_eval/logs/shengyuanhu__wmdp_unlearn_gd_ckpt_120_zephyr__wmdp_bio.log

@@ -0,0 +1,22 @@
+repo_id: shengyuanhu/wmdp_unlearn_gd_ckpt_120_zephyr
+split: bio
+snapshot: /zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/shengyuanhu__wmdp_unlearn_gd_ckpt_120_zephyr/models--shengyuanhu--wmdp_unlearn_gd_ckpt_120_zephyr/snapshots/b1cb0932176b4f37df99e108a5a97277af378914
+cmd: /zfspool/zangyihe/conda_envs/corgo/bin/python src/eval.py experiment=eval/wmdp/default model=zephyr-7b-beta model.model_args.pretrained_model_name_or_path=/zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/shengyuanhu__wmdp_unlearn_gd_ckpt_120_zephyr/models--shengyuanhu--wmdp_unlearn_gd_ckpt_120_zephyr/snapshots/b1cb0932176b4f37df99e108a5a97277af378914 model.tokenizer_args.pretrained_model_name_or_path=/zfspool/zangyihe/ulars/models/zephyr-7b-beta model.model_args.attn_implementation=sdpa model.model_args.torch_dtype=bfloat16 data_split=bio task_name=hf_wmdp_bio_shengyuanhu__wmdp_unlearn_gd_ckpt_120_zephyr paths.output_dir=/zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/shengyuanhu__wmdp_unlearn_gd_ckpt_120_zephyr__wmdp_bio eval.lm_eval.overwrite=true eval.lm_eval.tasks=[wmdp_bio] eval.lm_eval.simple_evaluate_args.batch_size=16
+
+/zfspool/zangyihe/conda_envs/corgo/lib/python3.11/site-packages/transformers/utils/hub.py:105: FutureWarning: Using `TRANSFORMERS_CACHE` is deprecated and will be removed in v5 of Transformers. Use `HF_HOME` instead.
+  warnings.warn(
+[2026-05-29 22:28:54,749] [INFO] [real_accelerator.py:219:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+
+[2026-05-29 22:29:20,879][lm_eval.models.huggingface][WARNING] - `pretrained` model kwarg is not of type `str`. Many other model arguments may be ignored. Please do not launch via accelerate or use `parallelize=True` if passing an existing model this way.
+[2026-05-29 22:29:20,895][lm_eval.models.huggingface][WARNING] - Passed an already-initialized model through `pretrained`, assuming single-process call to evaluate() or custom distributed integration
+[2026-05-29 22:29:20,896][evaluator][INFO] - ***** Running LMEval evaluation suite *****
+[2026-05-29 22:29:20,896][evaluator][INFO] - Fine-grained evaluations will be saved to: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/shengyuanhu__wmdp_unlearn_gd_ckpt_120_zephyr__wmdp_bio/LMEval_EVAL.json
+[2026-05-29 22:29:20,896][evaluator][INFO] - Aggregated evaluations will be summarised in: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/shengyuanhu__wmdp_unlearn_gd_ckpt_120_zephyr__wmdp_bio/LMEval_SUMMARY.json
+[2026-05-29 22:29:20,897][lm_eval.evaluator][INFO] - Setting random seed to 0 | Setting numpy seed to 1234 | Setting torch manual seed to 1234 | Setting fewshot manual seed to 1234
+[2026-05-29 22:29:20,897][lm_eval.evaluator][INFO] - Using pre-initialized model
+[2026-05-29 22:29:32,383][lm_eval.tasks][INFO] - Selected tasks:
+[2026-05-29 22:29:32,383][lm_eval.tasks][INFO] - Task: wmdp_bio (wmdp/wmdp_bio.yaml)
+[2026-05-29 22:29:32,384][lm_eval.api.task][INFO] - Building contexts for wmdp_bio on rank 0...
+
  0%|          | 0/1273 [00:00<?, ?it/s]
  8%|▊         | 96/1273 [00:00<00:01, 957.72it/s]
 15%|█▌        | 196/1273 [00:00<00:01, 976.81it/s]
 23%|██▎       | 296/1273 [00:00<00:00, 986.26it/s]
 31%|███       | 395/1273 [00:00<00:00, 983.12it/s]
 39%|███▉      | 495/1273 [00:00<00:00, 987.18it/s]
 47%|████▋     | 595/1273 [00:00<00:00, 990.76it/s]
 55%|█████▍    | 696/1273 [00:00<00:00, 994.73it/s]
 63%|██████▎   | 797/1273 [00:00<00:00, 996.80it/s]
 70%|███████   | 897/1273 [00:00<00:00, 995.34it/s]
 78%|███████▊  | 997/1273 [00:01<00:00, 992.88it/s]
 86%|████████▌ | 1097/1273 [00:01<00:00, 990.07it/s]
 94%|█████████▍| 1197/1273 [00:01<00:00, 990.07it/s]
+[2026-05-29 22:29:33,701][lm_eval.evaluator][INFO] - Running loglikelihood requests
+

data/ulars/supplement/standard_eval/logs/shengyuanhu__wmdp_unlearn_gd_ckpt_120_zephyr__wmdp_cyber.log

@@ -0,0 +1,22 @@
+repo_id: shengyuanhu/wmdp_unlearn_gd_ckpt_120_zephyr
+split: cyber
+snapshot: /zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/shengyuanhu__wmdp_unlearn_gd_ckpt_120_zephyr/models--shengyuanhu--wmdp_unlearn_gd_ckpt_120_zephyr/snapshots/b1cb0932176b4f37df99e108a5a97277af378914
+cmd: /zfspool/zangyihe/conda_envs/corgo/bin/python src/eval.py experiment=eval/wmdp/default model=zephyr-7b-beta model.model_args.pretrained_model_name_or_path=/zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/shengyuanhu__wmdp_unlearn_gd_ckpt_120_zephyr/models--shengyuanhu--wmdp_unlearn_gd_ckpt_120_zephyr/snapshots/b1cb0932176b4f37df99e108a5a97277af378914 model.tokenizer_args.pretrained_model_name_or_path=/zfspool/zangyihe/ulars/models/zephyr-7b-beta model.model_args.attn_implementation=sdpa model.model_args.torch_dtype=bfloat16 data_split=cyber task_name=hf_wmdp_cyber_shengyuanhu__wmdp_unlearn_gd_ckpt_120_zephyr paths.output_dir=/zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/shengyuanhu__wmdp_unlearn_gd_ckpt_120_zephyr__wmdp_cyber eval.lm_eval.overwrite=true eval.lm_eval.tasks=[wmdp_cyber] eval.lm_eval.simple_evaluate_args.batch_size=16
+
+/zfspool/zangyihe/conda_envs/corgo/lib/python3.11/site-packages/transformers/utils/hub.py:105: FutureWarning: Using `TRANSFORMERS_CACHE` is deprecated and will be removed in v5 of Transformers. Use `HF_HOME` instead.
+  warnings.warn(
+[2026-05-30 00:28:29,316] [INFO] [real_accelerator.py:219:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+
+[2026-05-30 00:28:44,044][lm_eval.models.huggingface][WARNING] - `pretrained` model kwarg is not of type `str`. Many other model arguments may be ignored. Please do not launch via accelerate or use `parallelize=True` if passing an existing model this way.
+[2026-05-30 00:28:44,059][lm_eval.models.huggingface][WARNING] - Passed an already-initialized model through `pretrained`, assuming single-process call to evaluate() or custom distributed integration
+[2026-05-30 00:28:44,060][evaluator][INFO] - ***** Running LMEval evaluation suite *****
+[2026-05-30 00:28:44,060][evaluator][INFO] - Fine-grained evaluations will be saved to: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/shengyuanhu__wmdp_unlearn_gd_ckpt_120_zephyr__wmdp_cyber/LMEval_EVAL.json
+[2026-05-30 00:28:44,060][evaluator][INFO] - Aggregated evaluations will be summarised in: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/shengyuanhu__wmdp_unlearn_gd_ckpt_120_zephyr__wmdp_cyber/LMEval_SUMMARY.json
+[2026-05-30 00:28:44,062][lm_eval.evaluator][INFO] - Setting random seed to 0 | Setting numpy seed to 1234 | Setting torch manual seed to 1234 | Setting fewshot manual seed to 1234
+[2026-05-30 00:28:44,062][lm_eval.evaluator][INFO] - Using pre-initialized model
+[2026-05-30 00:28:49,545][lm_eval.tasks][INFO] - Selected tasks:
+[2026-05-30 00:28:49,546][lm_eval.tasks][INFO] - Task: wmdp_cyber (wmdp/wmdp_cyber.yaml)
+[2026-05-30 00:28:49,546][lm_eval.api.task][INFO] - Building contexts for wmdp_cyber on rank 0...
+
  0%|          | 0/1987 [00:00<?, ?it/s]
  4%|▍         | 88/1987 [00:00<00:02, 877.48it/s]
  9%|▉         | 176/1987 [00:00<00:02, 774.37it/s]
 13%|█▎        | 255/1987 [00:00<00:02, 738.93it/s]
 17%|█▋        | 330/1987 [00:00<00:02, 721.11it/s]
 21%|██▏       | 425/1987 [00:00<00:01, 797.60it/s]
 26%|██▋       | 522/1987 [00:00<00:01, 853.28it/s]
 31%|███▏      | 622/1987 [00:00<00:01, 897.43it/s]
 36%|███▋      | 722/1987 [00:00<00:01, 928.75it/s]
 41%|████▏     | 822/1987 [00:00<00:01, 948.39it/s]
 46%|████▋     | 923/1987 [00:01<00:01, 964.70it/s]
 51%|█████▏    | 1021/1987 [00:01<00:00, 969.18it/s]
 56%|█████▋    | 1121/1987 [00:01<00:00, 977.55it/s]
 61%|██████▏   | 1222/1987 [00:01<00:00, 984.96it/s]
 67%|██████▋   | 1322/1987 [00:01<00:00, 988.87it/s]
 72%|███████▏  | 1422/1987 [00:01<00:00, 990.53it/s]
 77%|███████▋  | 1522/1987 [00:01<00:00, 990.54it/s]
 82%|████████▏ | 1622/1987 [00:01<00:00, 992.21it/s]
 87%|████████▋ | 1723/1987 [00:01<00:00, 994.87it/s]
 92%|█████████▏| 1823/1987 [00:01<00:00, 994.76it/s]
 97%|█████████▋| 1924/1987 [00:02<00:00, 996.61it/s]
+[2026-05-30 00:28:51,718][lm_eval.evaluator][INFO] - Running loglikelihood requests
+

data/ulars/supplement/standard_eval/logs/shengyuanhu__wmdp_unlearn_kl_ckpt_100_zephyr__wmdp_bio.log

@@ -0,0 +1,22 @@
+repo_id: shengyuanhu/wmdp_unlearn_kl_ckpt_100_zephyr
+split: bio
+snapshot: /zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/shengyuanhu__wmdp_unlearn_kl_ckpt_100_zephyr/models--shengyuanhu--wmdp_unlearn_kl_ckpt_100_zephyr/snapshots/c0b9cb75b53ffbf193136996257d18e994d967ad
+cmd: /zfspool/zangyihe/conda_envs/corgo/bin/python src/eval.py experiment=eval/wmdp/default model=zephyr-7b-beta model.model_args.pretrained_model_name_or_path=/zfspool/sharehub/models/hf_wmdp_zephyr_unlearned_cache/shengyuanhu__wmdp_unlearn_kl_ckpt_100_zephyr/models--shengyuanhu--wmdp_unlearn_kl_ckpt_100_zephyr/snapshots/c0b9cb75b53ffbf193136996257d18e994d967ad model.tokenizer_args.pretrained_model_name_or_path=/zfspool/zangyihe/ulars/models/zephyr-7b-beta model.model_args.attn_implementation=sdpa model.model_args.torch_dtype=bfloat16 data_split=bio task_name=hf_wmdp_bio_shengyuanhu__wmdp_unlearn_kl_ckpt_100_zephyr paths.output_dir=/zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/shengyuanhu__wmdp_unlearn_kl_ckpt_100_zephyr__wmdp_bio eval.lm_eval.overwrite=true eval.lm_eval.tasks=[wmdp_bio] eval.lm_eval.simple_evaluate_args.batch_size=16
+
+/zfspool/zangyihe/conda_envs/corgo/lib/python3.11/site-packages/transformers/utils/hub.py:105: FutureWarning: Using `TRANSFORMERS_CACHE` is deprecated and will be removed in v5 of Transformers. Use `HF_HOME` instead.
+  warnings.warn(
+[2026-05-30 00:53:11,140] [INFO] [real_accelerator.py:219:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+
+[2026-05-30 00:53:18,724][lm_eval.models.huggingface][WARNING] - `pretrained` model kwarg is not of type `str`. Many other model arguments may be ignored. Please do not launch via accelerate or use `parallelize=True` if passing an existing model this way.
+[2026-05-30 00:53:18,737][lm_eval.models.huggingface][WARNING] - Passed an already-initialized model through `pretrained`, assuming single-process call to evaluate() or custom distributed integration
+[2026-05-30 00:53:18,739][evaluator][INFO] - ***** Running LMEval evaluation suite *****
+[2026-05-30 00:53:18,739][evaluator][INFO] - Fine-grained evaluations will be saved to: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/shengyuanhu__wmdp_unlearn_kl_ckpt_100_zephyr__wmdp_bio/LMEval_EVAL.json
+[2026-05-30 00:53:18,739][evaluator][INFO] - Aggregated evaluations will be summarised in: /zfspool/zangyihe/ulars/hf-wmdp-eval/eval_outputs/shengyuanhu__wmdp_unlearn_kl_ckpt_100_zephyr__wmdp_bio/LMEval_SUMMARY.json
+[2026-05-30 00:53:18,740][lm_eval.evaluator][INFO] - Setting random seed to 0 | Setting numpy seed to 1234 | Setting torch manual seed to 1234 | Setting fewshot manual seed to 1234
+[2026-05-30 00:53:18,740][lm_eval.evaluator][INFO] - Using pre-initialized model
+[2026-05-30 00:53:22,349][lm_eval.tasks][INFO] - Selected tasks:
+[2026-05-30 00:53:22,349][lm_eval.tasks][INFO] - Task: wmdp_bio (wmdp/wmdp_bio.yaml)
+[2026-05-30 00:53:22,350][lm_eval.api.task][INFO] - Building contexts for wmdp_bio on rank 0...
+
  0%|          | 0/1273 [00:00<?, ?it/s]
  7%|▋         | 94/1273 [00:00<00:01, 935.76it/s]
 15%|█▌        | 194/1273 [00:00<00:01, 968.65it/s]
 23%|██▎       | 294/1273 [00:00<00:00, 980.89it/s]
 31%|███       | 393/1273 [00:00<00:00, 977.66it/s]
 39%|███▊      | 493/1273 [00:00<00:00, 984.77it/s]
 47%|████▋     | 594/1273 [00:00<00:00, 990.27it/s]
 55%|█████▍    | 695/1273 [00:00<00:00, 993.79it/s]
 63%|██████▎   | 796/1273 [00:00<00:00, 996.46it/s]
 70%|███████   | 897/1273 [00:00<00:00, 999.41it/s]
 78%|███████▊  | 998/1273 [00:01<00:00, 1000.23it/s]
 86%|████████▋ | 1099/1273 [00:01<00:00, 1001.13it/s]
 94%|█████████▍| 1200/1273 [00:01<00:00, 1002.01it/s]
+[2026-05-30 00:53:23,661][lm_eval.evaluator][INFO] - Running loglikelihood requests
+