deep-ignorance-unfiltered_unlearned_simnpo

This model was created by fine-tuning EleutherAI/deep-ignorance-unfiltered using the Simple NPO unlearning algorithm. The method is based on Meng et al. 2024. The goal of unlearning is to remove specific knowledge from a pretrained language model while preserving its general capabilities.

Hyperparameters

Parameter	Value
Base model	EleutherAI/deep-ignorance-unfiltered
Unlearning method	Simple NPO
Learning rate	3e-05
Epochs	3
Batch size	32
Max sequence length	2048
Optimizer	adamw
Gradient clipping	1.0
Gradient accumulation steps	1
Seed	42
W&B / run name	simnpo__ep3_lr3e-05_bs32_b0.01_rw1.0_mle2048_mli1024
Beta	0.01
Retain weight	1.0

Evaluation Results

Benchmark	Value
mmlu / acc	0.3796
mmlu / acc_stderr	0.0040
mmlu_abstract_algebra / acc	0.2800
mmlu_abstract_algebra / acc_stderr	0.0451
mmlu_anatomy / acc	0.4370
mmlu_anatomy / acc_stderr	0.0428
mmlu_astronomy / acc	0.3816
mmlu_astronomy / acc_stderr	0.0395
mmlu_business_ethics / acc	0.3900
mmlu_business_ethics / acc_stderr	0.0490
mmlu_clinical_knowledge / acc	0.3623
mmlu_clinical_knowledge / acc_stderr	0.0296
mmlu_college_biology / acc	0.4028
mmlu_college_biology / acc_stderr	0.0410
mmlu_college_chemistry / acc	0.2200
mmlu_college_chemistry / acc_stderr	0.0416
mmlu_college_computer_science / acc	0.2900
mmlu_college_computer_science / acc_stderr	0.0456
mmlu_college_mathematics / acc	0.2800
mmlu_college_mathematics / acc_stderr	0.0451
mmlu_college_medicine / acc	0.3584
mmlu_college_medicine / acc_stderr	0.0366
mmlu_college_physics / acc	0.1961
mmlu_college_physics / acc_stderr	0.0395
mmlu_computer_security / acc	0.5700
mmlu_computer_security / acc_stderr	0.0498
mmlu_conceptual_physics / acc	0.3872
mmlu_conceptual_physics / acc_stderr	0.0318
mmlu_econometrics / acc	0.2982
mmlu_econometrics / acc_stderr	0.0430
mmlu_electrical_engineering / acc	0.3655
mmlu_electrical_engineering / acc_stderr	0.0401
mmlu_elementary_mathematics / acc	0.2751
mmlu_elementary_mathematics / acc_stderr	0.0230
mmlu_formal_logic / acc	0.2143
mmlu_formal_logic / acc_stderr	0.0367
mmlu_global_facts / acc	0.3000
mmlu_global_facts / acc_stderr	0.0461
mmlu_high_school_biology / acc	0.3806
mmlu_high_school_biology / acc_stderr	0.0276
mmlu_high_school_chemistry / acc	0.3005
mmlu_high_school_chemistry / acc_stderr	0.0323
mmlu_high_school_computer_science / acc	0.4600
mmlu_high_school_computer_science / acc_stderr	0.0501
mmlu_high_school_european_history / acc	0.4303
mmlu_high_school_european_history / acc_stderr	0.0387
mmlu_high_school_geography / acc	0.4444
mmlu_high_school_geography / acc_stderr	0.0354
mmlu_high_school_government_and_politics / acc	0.5026
mmlu_high_school_government_and_politics / acc_stderr	0.0361
mmlu_high_school_macroeconomics / acc	0.3128
mmlu_high_school_macroeconomics / acc_stderr	0.0235
mmlu_high_school_mathematics / acc	0.2481
mmlu_high_school_mathematics / acc_stderr	0.0263
mmlu_high_school_microeconomics / acc	0.3739
mmlu_high_school_microeconomics / acc_stderr	0.0314
mmlu_high_school_physics / acc	0.3046
mmlu_high_school_physics / acc_stderr	0.0376
mmlu_high_school_psychology / acc	0.4771
mmlu_high_school_psychology / acc_stderr	0.0214
mmlu_high_school_statistics / acc	0.2315
mmlu_high_school_statistics / acc_stderr	0.0288
mmlu_high_school_us_history / acc	0.4314
mmlu_high_school_us_history / acc_stderr	0.0348
mmlu_high_school_world_history / acc	0.4937
mmlu_high_school_world_history / acc_stderr	0.0325
mmlu_human_aging / acc	0.4978
mmlu_human_aging / acc_stderr	0.0336
mmlu_human_sexuality / acc	0.4275
mmlu_human_sexuality / acc_stderr	0.0434
mmlu_humanities / acc	0.3571
mmlu_humanities / acc_stderr	0.0068
mmlu_international_law / acc	0.5537
mmlu_international_law / acc_stderr	0.0454
mmlu_jurisprudence / acc	0.4722
mmlu_jurisprudence / acc_stderr	0.0483
mmlu_logical_fallacies / acc	0.4233
mmlu_logical_fallacies / acc_stderr	0.0388
mmlu_machine_learning / acc	0.2946
mmlu_machine_learning / acc_stderr	0.0433
mmlu_management / acc	0.3883
mmlu_management / acc_stderr	0.0483
mmlu_marketing / acc	0.5812
mmlu_marketing / acc_stderr	0.0323
mmlu_medical_genetics / acc	0.4500
mmlu_medical_genetics / acc_stderr	0.0500
mmlu_miscellaneous / acc	0.5466
mmlu_miscellaneous / acc_stderr	0.0178
mmlu_moral_disputes / acc	0.4769
mmlu_moral_disputes / acc_stderr	0.0269
mmlu_moral_scenarios / acc	0.2514
mmlu_moral_scenarios / acc_stderr	0.0145
mmlu_nutrition / acc	0.3660
mmlu_nutrition / acc_stderr	0.0276
mmlu_other / acc	0.4281
mmlu_other / acc_stderr	0.0087
mmlu_philosophy / acc	0.4727
mmlu_philosophy / acc_stderr	0.0284
mmlu_prehistory / acc	0.4136
mmlu_prehistory / acc_stderr	0.0274
mmlu_professional_accounting / acc	0.3404
mmlu_professional_accounting / acc_stderr	0.0283
mmlu_professional_law / acc	0.2744
mmlu_professional_law / acc_stderr	0.0114
mmlu_professional_medicine / acc	0.2500
mmlu_professional_medicine / acc_stderr	0.0263
mmlu_professional_psychology / acc	0.3987
mmlu_professional_psychology / acc_stderr	0.0198
mmlu_public_relations / acc	0.4364
mmlu_public_relations / acc_stderr	0.0475
mmlu_security_studies / acc	0.3347
mmlu_security_studies / acc_stderr	0.0302
mmlu_social_sciences / acc	0.4202
mmlu_social_sciences / acc_stderr	0.0088
mmlu_sociology / acc	0.5721
mmlu_sociology / acc_stderr	0.0350
mmlu_stem / acc	0.3260
mmlu_stem / acc_stderr	0.0082
mmlu_us_foreign_policy / acc	0.5800
mmlu_us_foreign_policy / acc_stderr	0.0496
mmlu_virology / acc	0.4036
mmlu_virology / acc_stderr	0.0382
mmlu_world_religions / acc	0.5731
mmlu_world_religions / acc_stderr	0.0379
wikitext / bits_per_byte	1.0596
wikitext / bits_per_byte_stderr	N/A
wikitext / byte_perplexity	2.0844
wikitext / byte_perplexity_stderr	N/A
wikitext / word_perplexity	50.7864
wikitext / word_perplexity_stderr	N/A
wmdp_bio_categorized_mcqa / acc	0.4030
wmdp_bio_categorized_mcqa / acc_stderr	0.0136
wmdp_bio_cloze_verified / acc_norm	0.2695
wmdp_bio_cloze_verified / acc_norm_stderr	0.0135
wmdp_bio_robust / acc	0.3479
wmdp_bio_robust / acc_stderr	0.0162
wmdp_bio_robust_bioweapons_and_bioterrorism / acc	0.3263
wmdp_bio_robust_bioweapons_and_bioterrorism / acc_stderr	0.0341
wmdp_bio_robust_dual_use_virology / acc	0.3929
wmdp_bio_robust_dual_use_virology / acc_stderr	0.0940
wmdp_bio_robust_enhanced_potential_pandemic_pathogens / acc	0.2843
wmdp_bio_robust_enhanced_potential_pandemic_pathogens / acc_stderr	0.0449
wmdp_bio_robust_expanding_access_to_threat_vectors / acc	0.2857
wmdp_bio_robust_expanding_access_to_threat_vectors / acc_stderr	0.1010
wmdp_bio_robust_reverse_genetics_and_easy_editing / acc	0.3925
wmdp_bio_robust_reverse_genetics_and_easy_editing / acc_stderr	0.0359
wmdp_bio_robust_rewritten / acc	0.2322
wmdp_bio_robust_rewritten / acc_stderr	0.0086
wmdp_bio_robust_rewritten_gibberish / acc	0.2318
wmdp_bio_robust_rewritten_gibberish / acc_stderr	0.0148
wmdp_bio_robust_rewritten_nonsensical_biology / acc	0.2441
wmdp_bio_robust_rewritten_nonsensical_biology / acc_stderr	0.0151
wmdp_bio_robust_rewritten_real_words_sciency / acc	0.2207
wmdp_bio_robust_rewritten_real_words_sciency / acc_stderr	0.0146
wmdp_bio_robust_viral_vector_research / acc	0.3548
wmdp_bio_robust_viral_vector_research / acc_stderr	0.0259
wmdp_bio_shortcut / acc	0.5210
wmdp_bio_shortcut / acc_stderr	0.0249
wmdp_bio_shortcut_bioweapons_and_bioterrorism / acc	0.6383
wmdp_bio_shortcut_bioweapons_and_bioterrorism / acc_stderr	0.0708
wmdp_bio_shortcut_dual_use_virology / acc	0.4211
wmdp_bio_shortcut_dual_use_virology / acc_stderr	0.1164
wmdp_bio_shortcut_enhanced_potential_pandemic_pathogens / acc	0.5094
wmdp_bio_shortcut_enhanced_potential_pandemic_pathogens / acc_stderr	0.0693
wmdp_bio_shortcut_expanding_access_to_threat_vectors / acc	0.5556
wmdp_bio_shortcut_expanding_access_to_threat_vectors / acc_stderr	0.1757
wmdp_bio_shortcut_reverse_genetics_and_easy_editing / acc	0.5647
wmdp_bio_shortcut_reverse_genetics_and_easy_editing / acc_stderr	0.0541
wmdp_bio_shortcut_viral_vector_research / acc	0.4844
wmdp_bio_shortcut_viral_vector_research / acc_stderr	0.0362