{ "blimp_acc": 78.89, "blimp_n": 67000, "blimp_correct": 52859, "checkpoint_path": "/root/slayer149-balanced/runs/arcft-balanced/epoch-02.pt", "checkpoint_sha256": "79e4d06b1b95408c4b47e6d9fab34ed2a26609997b80d45698e6ec23db8724e8", "step": 1118, "training_tokens": null, "parameters": 149333081, "torch": "2.14.0+cu130", "gpu": "NVIDIA GeForce RTX 3090", "dataset_revisions": { "Salesforce/wikitext": "b08601e04326c79dfdd32d625aee71d232d685c3", "nyu-mll/blimp": "877fba0801ffb7cbd8c39c1ff314a46f053f6036", "allenai/ai2_arc": "210d026faf9955653af8916fad021475a3f00453" }, "recipe_sha256": "4a4ab81f9a749b04cf476672c0967512ec951a77947dea02fec3d77bf13fa091", "elapsed_seconds": 36.810004680999555, "protocol": "GLINT-derived pinned harness; CUDA BF16 forward, FP32 logprobs, context 256, no BOS, raw answer LL", "dataset_files": { "adjunct_island": [ "adjunct_island/train-00000-of-00001.parquet" ], "anaphor_gender_agreement": [ "anaphor_gender_agreement/train-00000-of-00001.parquet" ], "anaphor_number_agreement": [ "anaphor_number_agreement/train-00000-of-00001.parquet" ], "animate_subject_passive": [ "animate_subject_passive/train-00000-of-00001.parquet" ], "animate_subject_trans": [ "animate_subject_trans/train-00000-of-00001.parquet" ], "causative": [ "causative/train-00000-of-00001.parquet" ], "complex_NP_island": [ "complex_NP_island/train-00000-of-00001.parquet" ], "coordinate_structure_constraint_complex_left_branch": [ "coordinate_structure_constraint_complex_left_branch/train-00000-of-00001.parquet" ], "coordinate_structure_constraint_object_extraction": [ "coordinate_structure_constraint_object_extraction/train-00000-of-00001.parquet" ], "determiner_noun_agreement_1": [ "determiner_noun_agreement_1/train-00000-of-00001.parquet" ], "determiner_noun_agreement_2": [ "determiner_noun_agreement_2/train-00000-of-00001.parquet" ], "determiner_noun_agreement_irregular_1": [ "determiner_noun_agreement_irregular_1/train-00000-of-00001.parquet" ], "determiner_noun_agreement_irregular_2": [ "determiner_noun_agreement_irregular_2/train-00000-of-00001.parquet" ], "determiner_noun_agreement_with_adj_2": [ "determiner_noun_agreement_with_adj_2/train-00000-of-00001.parquet" ], "determiner_noun_agreement_with_adj_irregular_1": [ "determiner_noun_agreement_with_adj_irregular_1/train-00000-of-00001.parquet" ], "determiner_noun_agreement_with_adj_irregular_2": [ "determiner_noun_agreement_with_adj_irregular_2/train-00000-of-00001.parquet" ], "determiner_noun_agreement_with_adjective_1": [ "determiner_noun_agreement_with_adjective_1/train-00000-of-00001.parquet" ], "distractor_agreement_relational_noun": [ "distractor_agreement_relational_noun/train-00000-of-00001.parquet" ], "distractor_agreement_relative_clause": [ "distractor_agreement_relative_clause/train-00000-of-00001.parquet" ], "drop_argument": [ "drop_argument/train-00000-of-00001.parquet" ], "ellipsis_n_bar_1": [ "ellipsis_n_bar_1/train-00000-of-00001.parquet" ], "ellipsis_n_bar_2": [ "ellipsis_n_bar_2/train-00000-of-00001.parquet" ], "existential_there_object_raising": [ "existential_there_object_raising/train-00000-of-00001.parquet" ], "existential_there_quantifiers_1": [ "existential_there_quantifiers_1/train-00000-of-00001.parquet" ], "existential_there_quantifiers_2": [ "existential_there_quantifiers_2/train-00000-of-00001.parquet" ], "existential_there_subject_raising": [ "existential_there_subject_raising/train-00000-of-00001.parquet" ], "expletive_it_object_raising": [ "expletive_it_object_raising/train-00000-of-00001.parquet" ], "inchoative": [ "inchoative/train-00000-of-00001.parquet" ], "intransitive": [ "intransitive/train-00000-of-00001.parquet" ], "irregular_past_participle_adjectives": [ "irregular_past_participle_adjectives/train-00000-of-00001.parquet" ], "irregular_past_participle_verbs": [ "irregular_past_participle_verbs/train-00000-of-00001.parquet" ], "irregular_plural_subject_verb_agreement_1": [ "irregular_plural_subject_verb_agreement_1/train-00000-of-00001.parquet" ], "irregular_plural_subject_verb_agreement_2": [ "irregular_plural_subject_verb_agreement_2/train-00000-of-00001.parquet" ], "left_branch_island_echo_question": [ "left_branch_island_echo_question/train-00000-of-00001.parquet" ], "left_branch_island_simple_question": [ "left_branch_island_simple_question/train-00000-of-00001.parquet" ], "matrix_question_npi_licensor_present": [ "matrix_question_npi_licensor_present/train-00000-of-00001.parquet" ], "npi_present_1": [ "npi_present_1/train-00000-of-00001.parquet" ], "npi_present_2": [ "npi_present_2/train-00000-of-00001.parquet" ], "only_npi_licensor_present": [ "only_npi_licensor_present/train-00000-of-00001.parquet" ], "only_npi_scope": [ "only_npi_scope/train-00000-of-00001.parquet" ], "passive_1": [ "passive_1/train-00000-of-00001.parquet" ], "passive_2": [ "passive_2/train-00000-of-00001.parquet" ], "principle_A_c_command": [ "principle_A_c_command/train-00000-of-00001.parquet" ], "principle_A_case_1": [ "principle_A_case_1/train-00000-of-00001.parquet" ], "principle_A_case_2": [ "principle_A_case_2/train-00000-of-00001.parquet" ], "principle_A_domain_1": [ "principle_A_domain_1/train-00000-of-00001.parquet" ], "principle_A_domain_2": [ "principle_A_domain_2/train-00000-of-00001.parquet" ], "principle_A_domain_3": [ "principle_A_domain_3/train-00000-of-00001.parquet" ], "principle_A_reconstruction": [ "principle_A_reconstruction/train-00000-of-00001.parquet" ], "regular_plural_subject_verb_agreement_1": [ "regular_plural_subject_verb_agreement_1/train-00000-of-00001.parquet" ], "regular_plural_subject_verb_agreement_2": [ "regular_plural_subject_verb_agreement_2/train-00000-of-00001.parquet" ], "sentential_negation_npi_licensor_present": [ "sentential_negation_npi_licensor_present/train-00000-of-00001.parquet" ], "sentential_negation_npi_scope": [ "sentential_negation_npi_scope/train-00000-of-00001.parquet" ], "sentential_subject_island": [ "sentential_subject_island/train-00000-of-00001.parquet" ], "superlative_quantifiers_1": [ "superlative_quantifiers_1/train-00000-of-00001.parquet" ], "superlative_quantifiers_2": [ "superlative_quantifiers_2/train-00000-of-00001.parquet" ], "tough_vs_raising_1": [ "tough_vs_raising_1/train-00000-of-00001.parquet" ], "tough_vs_raising_2": [ "tough_vs_raising_2/train-00000-of-00001.parquet" ], "transitive": [ "transitive/train-00000-of-00001.parquet" ], "wh_island": [ "wh_island/train-00000-of-00001.parquet" ], "wh_questions_object_gap": [ "wh_questions_object_gap/train-00000-of-00001.parquet" ], "wh_questions_subject_gap": [ "wh_questions_subject_gap/train-00000-of-00001.parquet" ], "wh_questions_subject_gap_long_distance": [ "wh_questions_subject_gap_long_distance/train-00000-of-00001.parquet" ], "wh_vs_that_no_gap": [ "wh_vs_that_no_gap/train-00000-of-00001.parquet" ], "wh_vs_that_no_gap_long_distance": [ "wh_vs_that_no_gap_long_distance/train-00000-of-00001.parquet" ], "wh_vs_that_with_gap": [ "wh_vs_that_with_gap/train-00000-of-00001.parquet" ], "wh_vs_that_with_gap_long_distance": [ "wh_vs_that_with_gap_long_distance/train-00000-of-00001.parquet" ] } }