diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_abstract_algebra.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_abstract_algebra.yaml new file mode 100644 index 0000000000000000000000000000000000000000..76b96844f696cc22614cf2643858d4d6c9417754 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_abstract_algebra.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _lt_template_yaml +process_docs: !function utils.process_abstract_algebra +tag: global_mmlu_full_lt_stem_tasks +task: global_mmlu_full_lt_abstract_algebra diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_anatomy.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_anatomy.yaml new file mode 100644 index 0000000000000000000000000000000000000000..527c71074d96038c1414a9d24ac60c45b4dde78e --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_anatomy.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _lt_template_yaml +process_docs: !function utils.process_anatomy +tag: global_mmlu_full_lt_stem_tasks +task: global_mmlu_full_lt_anatomy diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_astronomy.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_astronomy.yaml new file mode 100644 index 0000000000000000000000000000000000000000..419b89e30ed8d5100d83d56a2d0be71888386e65 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_astronomy.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _lt_template_yaml +process_docs: !function utils.process_astronomy +tag: global_mmlu_full_lt_stem_tasks +task: global_mmlu_full_lt_astronomy diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_business_ethics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_business_ethics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..c51daa226b39284c5bf96ec531dc90fb1430fbd6 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_business_ethics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _lt_template_yaml +process_docs: !function utils.process_business_ethics +tag: global_mmlu_full_lt_other_tasks +task: global_mmlu_full_lt_business_ethics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_clinical_knowledge.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_clinical_knowledge.yaml new file mode 100644 index 0000000000000000000000000000000000000000..e023277406ff3c783c095e6996ac0c38211068bd --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_clinical_knowledge.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _lt_template_yaml +process_docs: !function utils.process_clinical_knowledge +tag: global_mmlu_full_lt_other_tasks +task: global_mmlu_full_lt_clinical_knowledge diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_college_biology.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_college_biology.yaml new file mode 100644 index 0000000000000000000000000000000000000000..c6fea6f851074e422ab06a10afee17138ee8c607 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_college_biology.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _lt_template_yaml +process_docs: !function utils.process_college_biology +tag: global_mmlu_full_lt_stem_tasks +task: global_mmlu_full_lt_college_biology diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_college_chemistry.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_college_chemistry.yaml new file mode 100644 index 0000000000000000000000000000000000000000..93b9a5610e16b589dd3e7e68f847490838e7f16d --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_college_chemistry.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _lt_template_yaml +process_docs: !function utils.process_college_chemistry +tag: global_mmlu_full_lt_stem_tasks +task: global_mmlu_full_lt_college_chemistry diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_college_computer_science.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_college_computer_science.yaml new file mode 100644 index 0000000000000000000000000000000000000000..8d0dcfdd9ad41771c5840124f3688e328b35b076 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_college_computer_science.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _lt_template_yaml +process_docs: !function utils.process_college_computer_science +tag: global_mmlu_full_lt_stem_tasks +task: global_mmlu_full_lt_college_computer_science diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_college_mathematics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_college_mathematics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..8d33b747d8276d3baa23b7187aaa5e2bfc9f7939 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_college_mathematics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _lt_template_yaml +process_docs: !function utils.process_college_mathematics +tag: global_mmlu_full_lt_stem_tasks +task: global_mmlu_full_lt_college_mathematics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_college_medicine.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_college_medicine.yaml new file mode 100644 index 0000000000000000000000000000000000000000..ad74dbb15fcfdbc529e9e0440bf2122ee0aa9e8a --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_college_medicine.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _lt_template_yaml +process_docs: !function utils.process_college_medicine +tag: global_mmlu_full_lt_other_tasks +task: global_mmlu_full_lt_college_medicine diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_college_physics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_college_physics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..3c69754b32b3d7affd9a2313a8b2be9f40dbbaf2 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_college_physics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _lt_template_yaml +process_docs: !function utils.process_college_physics +tag: global_mmlu_full_lt_stem_tasks +task: global_mmlu_full_lt_college_physics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_computer_security.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_computer_security.yaml new file mode 100644 index 0000000000000000000000000000000000000000..d78f3a545da42447e590ef4f0156e6ae49f505d1 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_computer_security.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _lt_template_yaml +process_docs: !function utils.process_computer_security +tag: global_mmlu_full_lt_stem_tasks +task: global_mmlu_full_lt_computer_security diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_conceptual_physics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_conceptual_physics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..3e7b5e49e6e2028dd176336874c902ea68fccb5f --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_conceptual_physics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _lt_template_yaml +process_docs: !function utils.process_conceptual_physics +tag: global_mmlu_full_lt_stem_tasks +task: global_mmlu_full_lt_conceptual_physics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_econometrics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_econometrics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..6d0085ab968a802e03eb40c8526d8d249d382ce6 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_econometrics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _lt_template_yaml +process_docs: !function utils.process_econometrics +tag: global_mmlu_full_lt_social_sciences_tasks +task: global_mmlu_full_lt_econometrics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_electrical_engineering.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_electrical_engineering.yaml new file mode 100644 index 0000000000000000000000000000000000000000..284dfe9e27a30ed0e529e80d004c3f273bb2b575 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_electrical_engineering.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _lt_template_yaml +process_docs: !function utils.process_electrical_engineering +tag: global_mmlu_full_lt_stem_tasks +task: global_mmlu_full_lt_electrical_engineering diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_elementary_mathematics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_elementary_mathematics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..7e9a01030ee326f7a950704b589c8169beea6a67 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_elementary_mathematics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _lt_template_yaml +process_docs: !function utils.process_elementary_mathematics +tag: global_mmlu_full_lt_stem_tasks +task: global_mmlu_full_lt_elementary_mathematics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_formal_logic.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_formal_logic.yaml new file mode 100644 index 0000000000000000000000000000000000000000..ec9a665b52d2a0c8aa1af6787eff4cb839c998ee --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_formal_logic.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _lt_template_yaml +process_docs: !function utils.process_formal_logic +tag: global_mmlu_full_lt_humanities_tasks +task: global_mmlu_full_lt_formal_logic diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_global_facts.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_global_facts.yaml new file mode 100644 index 0000000000000000000000000000000000000000..d81a9470486926fd53c9cc4f9d2d8d8a8fc19eb1 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_global_facts.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _lt_template_yaml +process_docs: !function utils.process_global_facts +tag: global_mmlu_full_lt_other_tasks +task: global_mmlu_full_lt_global_facts diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_high_school_biology.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_high_school_biology.yaml new file mode 100644 index 0000000000000000000000000000000000000000..139376cc770a8a0c321ff3241147afff31e5b164 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_high_school_biology.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _lt_template_yaml +process_docs: !function utils.process_high_school_biology +tag: global_mmlu_full_lt_stem_tasks +task: global_mmlu_full_lt_high_school_biology diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_high_school_chemistry.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_high_school_chemistry.yaml new file mode 100644 index 0000000000000000000000000000000000000000..87112d8c77823c53196ae5f1ec70e5d43cd9cfee --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_high_school_chemistry.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _lt_template_yaml +process_docs: !function utils.process_high_school_chemistry +tag: global_mmlu_full_lt_stem_tasks +task: global_mmlu_full_lt_high_school_chemistry diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_high_school_computer_science.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_high_school_computer_science.yaml new file mode 100644 index 0000000000000000000000000000000000000000..2324bb288496581d56df81e035f2a0faf16e0040 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_high_school_computer_science.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _lt_template_yaml +process_docs: !function utils.process_high_school_computer_science +tag: global_mmlu_full_lt_stem_tasks +task: global_mmlu_full_lt_high_school_computer_science diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_high_school_european_history.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_high_school_european_history.yaml new file mode 100644 index 0000000000000000000000000000000000000000..5f365fab1d50d5cc33935198ab2f5e3d0a80d500 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_high_school_european_history.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _lt_template_yaml +process_docs: !function utils.process_high_school_european_history +tag: global_mmlu_full_lt_humanities_tasks +task: global_mmlu_full_lt_high_school_european_history diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_high_school_geography.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_high_school_geography.yaml new file mode 100644 index 0000000000000000000000000000000000000000..e3a6f9214771c7281150b390193cad51fe844c98 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_high_school_geography.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _lt_template_yaml +process_docs: !function utils.process_high_school_geography +tag: global_mmlu_full_lt_social_sciences_tasks +task: global_mmlu_full_lt_high_school_geography diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_high_school_government_and_politics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_high_school_government_and_politics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..526b68ed9051c96d53fe8c54f95c386173ccaacb --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_high_school_government_and_politics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _lt_template_yaml +process_docs: !function utils.process_high_school_government_and_politics +tag: global_mmlu_full_lt_social_sciences_tasks +task: global_mmlu_full_lt_high_school_government_and_politics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_high_school_macroeconomics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_high_school_macroeconomics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..e14b1dcea2da09877e98827770e1b14fe7b55cb2 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_high_school_macroeconomics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _lt_template_yaml +process_docs: !function utils.process_high_school_macroeconomics +tag: global_mmlu_full_lt_social_sciences_tasks +task: global_mmlu_full_lt_high_school_macroeconomics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_high_school_mathematics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_high_school_mathematics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..1cdf5c90b9cde143640c51d349fb23fef3f0ada7 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_high_school_mathematics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _lt_template_yaml +process_docs: !function utils.process_high_school_mathematics +tag: global_mmlu_full_lt_stem_tasks +task: global_mmlu_full_lt_high_school_mathematics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_high_school_microeconomics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_high_school_microeconomics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..a2f2a2103bf5df058fbbb5540b83ec9b5216a955 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_high_school_microeconomics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _lt_template_yaml +process_docs: !function utils.process_high_school_microeconomics +tag: global_mmlu_full_lt_social_sciences_tasks +task: global_mmlu_full_lt_high_school_microeconomics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_high_school_physics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_high_school_physics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..bd363709173087d90cfbb9e65cf1df80361f74af --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_high_school_physics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _lt_template_yaml +process_docs: !function utils.process_high_school_physics +tag: global_mmlu_full_lt_stem_tasks +task: global_mmlu_full_lt_high_school_physics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_high_school_psychology.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_high_school_psychology.yaml new file mode 100644 index 0000000000000000000000000000000000000000..aad65a13643885afa6d00742737983ff2376b4c0 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_high_school_psychology.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _lt_template_yaml +process_docs: !function utils.process_high_school_psychology +tag: global_mmlu_full_lt_social_sciences_tasks +task: global_mmlu_full_lt_high_school_psychology diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_high_school_statistics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_high_school_statistics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..6dd6d699a1a2eee445ba12047e4687c86f7e8440 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_high_school_statistics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _lt_template_yaml +process_docs: !function utils.process_high_school_statistics +tag: global_mmlu_full_lt_stem_tasks +task: global_mmlu_full_lt_high_school_statistics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_high_school_us_history.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_high_school_us_history.yaml new file mode 100644 index 0000000000000000000000000000000000000000..5fb0ee1e513d860418cb68acb630cfc7ef1aa10d --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_high_school_us_history.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _lt_template_yaml +process_docs: !function utils.process_high_school_us_history +tag: global_mmlu_full_lt_humanities_tasks +task: global_mmlu_full_lt_high_school_us_history diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_high_school_world_history.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_high_school_world_history.yaml new file mode 100644 index 0000000000000000000000000000000000000000..75f2769ada0253604eec937509d7927532b0c377 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_high_school_world_history.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _lt_template_yaml +process_docs: !function utils.process_high_school_world_history +tag: global_mmlu_full_lt_humanities_tasks +task: global_mmlu_full_lt_high_school_world_history diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_human_aging.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_human_aging.yaml new file mode 100644 index 0000000000000000000000000000000000000000..beb27e9bd2471b6de27150e60d89df8fd6103d27 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_human_aging.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _lt_template_yaml +process_docs: !function utils.process_human_aging +tag: global_mmlu_full_lt_other_tasks +task: global_mmlu_full_lt_human_aging diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_human_sexuality.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_human_sexuality.yaml new file mode 100644 index 0000000000000000000000000000000000000000..c9d952c39b0a13ff0887fdb671535776babe9088 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_human_sexuality.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _lt_template_yaml +process_docs: !function utils.process_human_sexuality +tag: global_mmlu_full_lt_social_sciences_tasks +task: global_mmlu_full_lt_human_sexuality diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_international_law.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_international_law.yaml new file mode 100644 index 0000000000000000000000000000000000000000..f77adf9b399d0dfbf28debccb427d95ab2467245 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_international_law.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _lt_template_yaml +process_docs: !function utils.process_international_law +tag: global_mmlu_full_lt_humanities_tasks +task: global_mmlu_full_lt_international_law diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_jurisprudence.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_jurisprudence.yaml new file mode 100644 index 0000000000000000000000000000000000000000..e6be84fde4ca8de4bc8d1795ad8ba6a223bb3d50 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_jurisprudence.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _lt_template_yaml +process_docs: !function utils.process_jurisprudence +tag: global_mmlu_full_lt_humanities_tasks +task: global_mmlu_full_lt_jurisprudence diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_logical_fallacies.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_logical_fallacies.yaml new file mode 100644 index 0000000000000000000000000000000000000000..ad597b27aa2717ae8aeec74c69a1595bfafe129a --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_logical_fallacies.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _lt_template_yaml +process_docs: !function utils.process_logical_fallacies +tag: global_mmlu_full_lt_humanities_tasks +task: global_mmlu_full_lt_logical_fallacies diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_machine_learning.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_machine_learning.yaml new file mode 100644 index 0000000000000000000000000000000000000000..eb06a871b072676ed1e4d4e5c09faa9657346a22 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_machine_learning.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _lt_template_yaml +process_docs: !function utils.process_machine_learning +tag: global_mmlu_full_lt_stem_tasks +task: global_mmlu_full_lt_machine_learning diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_management.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_management.yaml new file mode 100644 index 0000000000000000000000000000000000000000..e1885ad3faa748e02b78cd1691e8cf4f752da223 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_management.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _lt_template_yaml +process_docs: !function utils.process_management +tag: global_mmlu_full_lt_other_tasks +task: global_mmlu_full_lt_management diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_marketing.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_marketing.yaml new file mode 100644 index 0000000000000000000000000000000000000000..2dc83089094d69a5872cd596374c7ef9d3707ac2 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_marketing.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _lt_template_yaml +process_docs: !function utils.process_marketing +tag: global_mmlu_full_lt_other_tasks +task: global_mmlu_full_lt_marketing diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_medical_genetics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_medical_genetics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..b67d321e40986fd04d93a75073868e98c7b0998f --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_medical_genetics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _lt_template_yaml +process_docs: !function utils.process_medical_genetics +tag: global_mmlu_full_lt_other_tasks +task: global_mmlu_full_lt_medical_genetics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_miscellaneous.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_miscellaneous.yaml new file mode 100644 index 0000000000000000000000000000000000000000..2c744613c6e76867f77f901428b119071474c06d --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_miscellaneous.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _lt_template_yaml +process_docs: !function utils.process_miscellaneous +tag: global_mmlu_full_lt_other_tasks +task: global_mmlu_full_lt_miscellaneous diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_moral_disputes.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_moral_disputes.yaml new file mode 100644 index 0000000000000000000000000000000000000000..09e6f044fcf2202fcc6d8278649996613662b261 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_moral_disputes.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _lt_template_yaml +process_docs: !function utils.process_moral_disputes +tag: global_mmlu_full_lt_humanities_tasks +task: global_mmlu_full_lt_moral_disputes diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_moral_scenarios.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_moral_scenarios.yaml new file mode 100644 index 0000000000000000000000000000000000000000..bb8dd3307fa10a0c5628413331545e7773c4d245 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_moral_scenarios.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _lt_template_yaml +process_docs: !function utils.process_moral_scenarios +tag: global_mmlu_full_lt_humanities_tasks +task: global_mmlu_full_lt_moral_scenarios diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_nutrition.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_nutrition.yaml new file mode 100644 index 0000000000000000000000000000000000000000..0b1a85566793f56997423ae6063b75b44aa017d7 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_nutrition.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _lt_template_yaml +process_docs: !function utils.process_nutrition +tag: global_mmlu_full_lt_other_tasks +task: global_mmlu_full_lt_nutrition diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_philosophy.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_philosophy.yaml new file mode 100644 index 0000000000000000000000000000000000000000..aab1d556c86ddddcc47ec278c45c6dff556641ad --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_philosophy.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _lt_template_yaml +process_docs: !function utils.process_philosophy +tag: global_mmlu_full_lt_humanities_tasks +task: global_mmlu_full_lt_philosophy diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_prehistory.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_prehistory.yaml new file mode 100644 index 0000000000000000000000000000000000000000..ac93dd6cc9c14abe297b765b04a71c8b51a44ab4 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_prehistory.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _lt_template_yaml +process_docs: !function utils.process_prehistory +tag: global_mmlu_full_lt_humanities_tasks +task: global_mmlu_full_lt_prehistory diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_professional_accounting.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_professional_accounting.yaml new file mode 100644 index 0000000000000000000000000000000000000000..6be78ec390713a42935600ead6cdf9734c8f5cd4 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_professional_accounting.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _lt_template_yaml +process_docs: !function utils.process_professional_accounting +tag: global_mmlu_full_lt_other_tasks +task: global_mmlu_full_lt_professional_accounting diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_professional_law.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_professional_law.yaml new file mode 100644 index 0000000000000000000000000000000000000000..60b6cdcc0b9016e13ed117af0ff9f38728f536cc --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_professional_law.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _lt_template_yaml +process_docs: !function utils.process_professional_law +tag: global_mmlu_full_lt_humanities_tasks +task: global_mmlu_full_lt_professional_law diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_professional_medicine.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_professional_medicine.yaml new file mode 100644 index 0000000000000000000000000000000000000000..dd899676e6dfedcda15b8d30b048906b81983604 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_professional_medicine.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _lt_template_yaml +process_docs: !function utils.process_professional_medicine +tag: global_mmlu_full_lt_other_tasks +task: global_mmlu_full_lt_professional_medicine diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_professional_psychology.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_professional_psychology.yaml new file mode 100644 index 0000000000000000000000000000000000000000..bd796e8b55c413a9c1937b92ad797174e8a0186b --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_professional_psychology.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _lt_template_yaml +process_docs: !function utils.process_professional_psychology +tag: global_mmlu_full_lt_social_sciences_tasks +task: global_mmlu_full_lt_professional_psychology diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_public_relations.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_public_relations.yaml new file mode 100644 index 0000000000000000000000000000000000000000..3c6e5f397293a0cda432923d9c9bb8d71b305af3 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_public_relations.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _lt_template_yaml +process_docs: !function utils.process_public_relations +tag: global_mmlu_full_lt_social_sciences_tasks +task: global_mmlu_full_lt_public_relations diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_security_studies.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_security_studies.yaml new file mode 100644 index 0000000000000000000000000000000000000000..9eb9957d2df677f8d81aa8cd4123671cef29c53e --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_security_studies.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _lt_template_yaml +process_docs: !function utils.process_security_studies +tag: global_mmlu_full_lt_social_sciences_tasks +task: global_mmlu_full_lt_security_studies diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_sociology.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_sociology.yaml new file mode 100644 index 0000000000000000000000000000000000000000..2e17f95a66b8d28a0b68e6050241d6140988b7d8 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_sociology.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _lt_template_yaml +process_docs: !function utils.process_sociology +tag: global_mmlu_full_lt_social_sciences_tasks +task: global_mmlu_full_lt_sociology diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_us_foreign_policy.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_us_foreign_policy.yaml new file mode 100644 index 0000000000000000000000000000000000000000..d39bb63caec03b27a46a16c3094aacd15eb4d567 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_us_foreign_policy.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _lt_template_yaml +process_docs: !function utils.process_us_foreign_policy +tag: global_mmlu_full_lt_social_sciences_tasks +task: global_mmlu_full_lt_us_foreign_policy diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_virology.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_virology.yaml new file mode 100644 index 0000000000000000000000000000000000000000..b8482a618bda42c8917f010a31bc259dbb4386e2 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_virology.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _lt_template_yaml +process_docs: !function utils.process_virology +tag: global_mmlu_full_lt_other_tasks +task: global_mmlu_full_lt_virology diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_world_religions.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_world_religions.yaml new file mode 100644 index 0000000000000000000000000000000000000000..a86af60dce82c39b95efc1da3e430334d30c1df6 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/global_mmlu_full_lt_world_religions.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _lt_template_yaml +process_docs: !function utils.process_world_religions +tag: global_mmlu_full_lt_humanities_tasks +task: global_mmlu_full_lt_world_religions diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/utils.py b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/utils.py new file mode 100644 index 0000000000000000000000000000000000000000..7df72cb061f0fecba46e15ff9b57552817979afb --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/lt/utils.py @@ -0,0 +1,73 @@ +from functools import partial + + +SUBJECTS = [ + "abstract_algebra", + "anatomy", + "astronomy", + "business_ethics", + "clinical_knowledge", + "college_biology", + "college_chemistry", + "college_computer_science", + "college_mathematics", + "college_medicine", + "college_physics", + "computer_security", + "conceptual_physics", + "econometrics", + "electrical_engineering", + "elementary_mathematics", + "formal_logic", + "global_facts", + "high_school_biology", + "high_school_chemistry", + "high_school_computer_science", + "high_school_european_history", + "high_school_geography", + "high_school_government_and_politics", + "high_school_macroeconomics", + "high_school_mathematics", + "high_school_microeconomics", + "high_school_physics", + "high_school_psychology", + "high_school_statistics", + "high_school_us_history", + "high_school_world_history", + "human_aging", + "human_sexuality", + "international_law", + "jurisprudence", + "logical_fallacies", + "machine_learning", + "management", + "marketing", + "medical_genetics", + "miscellaneous", + "moral_disputes", + "moral_scenarios", + "nutrition", + "philosophy", + "prehistory", + "professional_accounting", + "professional_law", + "professional_medicine", + "professional_psychology", + "public_relations", + "security_studies", + "sociology", + "us_foreign_policy", + "virology", + "world_religions", +] + + +def process_docs(dataset, subject): + return dataset.filter(lambda x: x["subject"] == subject) + + +process_functions = { + f"process_{subject}": partial(process_docs, subject=subject) for subject in SUBJECTS +} + +globals().update(process_functions) diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/_global_mmlu_full_mg.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/_global_mmlu_full_mg.yaml new file mode 100644 index 0000000000000000000000000000000000000000..bcc6cd81adfab398e1164a063d6303218bbea53f --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/_global_mmlu_full_mg.yaml @@ -0,0 +1,11 @@ +group: global_mmlu_full_mg +task: + - global_mmlu_full_mg_stem + - global_mmlu_full_mg_other + - global_mmlu_full_mg_social_sciences + - global_mmlu_full_mg_humanities +aggregate_metric_list: + - metric: acc + weight_by_size: True +metadata: + version: 0.0 diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/_global_mmlu_full_mg_humanities.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/_global_mmlu_full_mg_humanities.yaml new file mode 100644 index 0000000000000000000000000000000000000000..76b08f6c776b956f312b05f879ba0a278025607f --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/_global_mmlu_full_mg_humanities.yaml @@ -0,0 +1,8 @@ +group: global_mmlu_full_mg_humanities +task: + - global_mmlu_full_mg_humanities_tasks +aggregate_metric_list: + - metric: acc + weight_by_size: true +metadata: + version: 0.0 diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/_global_mmlu_full_mg_other.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/_global_mmlu_full_mg_other.yaml new file mode 100644 index 0000000000000000000000000000000000000000..0006af4c995b0abfd5d34b87e363218b2a445c20 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/_global_mmlu_full_mg_other.yaml @@ -0,0 +1,8 @@ +group: global_mmlu_full_mg_other +task: + - global_mmlu_full_mg_other_tasks +aggregate_metric_list: + - metric: acc + weight_by_size: true +metadata: + version: 0.0 diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/_global_mmlu_full_mg_social_sciences.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/_global_mmlu_full_mg_social_sciences.yaml new file mode 100644 index 0000000000000000000000000000000000000000..9cfe4f5b02b6fb983c3ccb65f9a4345fa3b39222 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/_global_mmlu_full_mg_social_sciences.yaml @@ -0,0 +1,8 @@ +group: global_mmlu_full_mg_social_sciences +task: + - global_mmlu_full_mg_social_sciences_tasks +aggregate_metric_list: + - metric: acc + weight_by_size: true +metadata: + version: 0.0 diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/_global_mmlu_full_mg_stem.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/_global_mmlu_full_mg_stem.yaml new file mode 100644 index 0000000000000000000000000000000000000000..bdc719d1076ec7081b6547d655887bfb6c81dbc0 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/_global_mmlu_full_mg_stem.yaml @@ -0,0 +1,8 @@ +group: global_mmlu_full_mg_stem +task: + - global_mmlu_full_mg_stem_tasks +aggregate_metric_list: + - metric: acc + weight_by_size: true +metadata: + version: 0.0 diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/_mg_template_yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/_mg_template_yaml new file mode 100644 index 0000000000000000000000000000000000000000..4aa97b27f3b6926ba42070c81c64b004e8f331ac --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/_mg_template_yaml @@ -0,0 +1,16 @@ +dataset_path: CohereForAI/Global-MMLU +dataset_name: mg +test_split: test +fewshot_split: dev +fewshot_config: + sampler: first_n +output_type: multiple_choice +doc_to_text: "{{question.strip()}}\nA. {{option_a}}\nB. {{option_b}}\nC. {{option_c}}\nD. {{option_d}}\nAnswer:" +doc_to_choice: ["A", "B", "C", "D"] +doc_to_target: answer +metric_list: + - metric: acc + aggregation: mean + higher_is_better: true +metadata: + version: 0.0 diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_abstract_algebra.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_abstract_algebra.yaml new file mode 100644 index 0000000000000000000000000000000000000000..bea850ed14165bc4e32806c9828f2074b86ac36f --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_abstract_algebra.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _mg_template_yaml +process_docs: !function utils.process_abstract_algebra +tag: global_mmlu_full_mg_stem_tasks +task: global_mmlu_full_mg_abstract_algebra diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_anatomy.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_anatomy.yaml new file mode 100644 index 0000000000000000000000000000000000000000..1cf6c1169aa1379edd85737d0d4ac21875471679 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_anatomy.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _mg_template_yaml +process_docs: !function utils.process_anatomy +tag: global_mmlu_full_mg_stem_tasks +task: global_mmlu_full_mg_anatomy diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_astronomy.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_astronomy.yaml new file mode 100644 index 0000000000000000000000000000000000000000..df582b2720bdcb16d660a8f4c8c1aa7bca0bf76d --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_astronomy.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _mg_template_yaml +process_docs: !function utils.process_astronomy +tag: global_mmlu_full_mg_stem_tasks +task: global_mmlu_full_mg_astronomy diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_business_ethics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_business_ethics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..a63513426992fc2a5a1a3004fab4fdea268a84e5 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_business_ethics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _mg_template_yaml +process_docs: !function utils.process_business_ethics +tag: global_mmlu_full_mg_other_tasks +task: global_mmlu_full_mg_business_ethics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_clinical_knowledge.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_clinical_knowledge.yaml new file mode 100644 index 0000000000000000000000000000000000000000..21003af5044b9535e107b4fa357c880dab55a8b6 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_clinical_knowledge.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _mg_template_yaml +process_docs: !function utils.process_clinical_knowledge +tag: global_mmlu_full_mg_other_tasks +task: global_mmlu_full_mg_clinical_knowledge diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_college_biology.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_college_biology.yaml new file mode 100644 index 0000000000000000000000000000000000000000..d305ca94e852fcc68c1541a2f0b31174c7871dd9 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_college_biology.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _mg_template_yaml +process_docs: !function utils.process_college_biology +tag: global_mmlu_full_mg_stem_tasks +task: global_mmlu_full_mg_college_biology diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_college_chemistry.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_college_chemistry.yaml new file mode 100644 index 0000000000000000000000000000000000000000..7ccaffb9dce404179a4660422a10a0dc9ab065a0 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_college_chemistry.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _mg_template_yaml +process_docs: !function utils.process_college_chemistry +tag: global_mmlu_full_mg_stem_tasks +task: global_mmlu_full_mg_college_chemistry diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_college_computer_science.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_college_computer_science.yaml new file mode 100644 index 0000000000000000000000000000000000000000..248f72c3df3a11dc885ea56bf87ba09a9b90fd0d --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_college_computer_science.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _mg_template_yaml +process_docs: !function utils.process_college_computer_science +tag: global_mmlu_full_mg_stem_tasks +task: global_mmlu_full_mg_college_computer_science diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_college_mathematics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_college_mathematics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..fb817aaeba681738f393148349190a197751d3a1 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_college_mathematics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _mg_template_yaml +process_docs: !function utils.process_college_mathematics +tag: global_mmlu_full_mg_stem_tasks +task: global_mmlu_full_mg_college_mathematics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_college_medicine.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_college_medicine.yaml new file mode 100644 index 0000000000000000000000000000000000000000..4fdbee03bfff14a701eeebab940dcdebee55dd44 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_college_medicine.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _mg_template_yaml +process_docs: !function utils.process_college_medicine +tag: global_mmlu_full_mg_other_tasks +task: global_mmlu_full_mg_college_medicine diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_college_physics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_college_physics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..493bdf87b557a384a093d779c2e43d736f3f793c --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_college_physics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _mg_template_yaml +process_docs: !function utils.process_college_physics +tag: global_mmlu_full_mg_stem_tasks +task: global_mmlu_full_mg_college_physics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_computer_security.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_computer_security.yaml new file mode 100644 index 0000000000000000000000000000000000000000..c5054eb639c576647380db82e046d6c814c63212 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_computer_security.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _mg_template_yaml +process_docs: !function utils.process_computer_security +tag: global_mmlu_full_mg_stem_tasks +task: global_mmlu_full_mg_computer_security diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_conceptual_physics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_conceptual_physics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..44a13a70eabbdb30b07de7844746c9af79e9bfe1 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_conceptual_physics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _mg_template_yaml +process_docs: !function utils.process_conceptual_physics +tag: global_mmlu_full_mg_stem_tasks +task: global_mmlu_full_mg_conceptual_physics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_econometrics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_econometrics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..2c5d029adc532558620e77d7f4f5c35205f783a1 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_econometrics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _mg_template_yaml +process_docs: !function utils.process_econometrics +tag: global_mmlu_full_mg_social_sciences_tasks +task: global_mmlu_full_mg_econometrics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_electrical_engineering.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_electrical_engineering.yaml new file mode 100644 index 0000000000000000000000000000000000000000..1e5ece337159462750cd43a61f0f0ae0fcf8e5f6 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_electrical_engineering.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _mg_template_yaml +process_docs: !function utils.process_electrical_engineering +tag: global_mmlu_full_mg_stem_tasks +task: global_mmlu_full_mg_electrical_engineering diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_elementary_mathematics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_elementary_mathematics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..4d62c7585ba2c757bd143f057ab7a6e59c92c03e --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_elementary_mathematics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _mg_template_yaml +process_docs: !function utils.process_elementary_mathematics +tag: global_mmlu_full_mg_stem_tasks +task: global_mmlu_full_mg_elementary_mathematics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_formal_logic.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_formal_logic.yaml new file mode 100644 index 0000000000000000000000000000000000000000..e5dc67d05b6978c5515d857638f07fcd5ebc2696 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_formal_logic.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _mg_template_yaml +process_docs: !function utils.process_formal_logic +tag: global_mmlu_full_mg_humanities_tasks +task: global_mmlu_full_mg_formal_logic diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_global_facts.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_global_facts.yaml new file mode 100644 index 0000000000000000000000000000000000000000..2712e9b4691d1c64c5a45713232d9cbcd63b0837 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_global_facts.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _mg_template_yaml +process_docs: !function utils.process_global_facts +tag: global_mmlu_full_mg_other_tasks +task: global_mmlu_full_mg_global_facts diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_high_school_biology.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_high_school_biology.yaml new file mode 100644 index 0000000000000000000000000000000000000000..c58957e03ccd6c85d76e1ddf32ef875bf2f4065c --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_high_school_biology.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _mg_template_yaml +process_docs: !function utils.process_high_school_biology +tag: global_mmlu_full_mg_stem_tasks +task: global_mmlu_full_mg_high_school_biology diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_high_school_chemistry.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_high_school_chemistry.yaml new file mode 100644 index 0000000000000000000000000000000000000000..707b73564049e39a10ff7e0b87dcc128c36c1f2e --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_high_school_chemistry.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _mg_template_yaml +process_docs: !function utils.process_high_school_chemistry +tag: global_mmlu_full_mg_stem_tasks +task: global_mmlu_full_mg_high_school_chemistry diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_high_school_computer_science.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_high_school_computer_science.yaml new file mode 100644 index 0000000000000000000000000000000000000000..d7afd5a7425e7161ff8c2b3d61f9abd77658ff9f --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_high_school_computer_science.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _mg_template_yaml +process_docs: !function utils.process_high_school_computer_science +tag: global_mmlu_full_mg_stem_tasks +task: global_mmlu_full_mg_high_school_computer_science diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_high_school_european_history.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_high_school_european_history.yaml new file mode 100644 index 0000000000000000000000000000000000000000..b6391ee4473db316dbd5685e3e94020648a4c6c9 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_high_school_european_history.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _mg_template_yaml +process_docs: !function utils.process_high_school_european_history +tag: global_mmlu_full_mg_humanities_tasks +task: global_mmlu_full_mg_high_school_european_history diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_high_school_geography.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_high_school_geography.yaml new file mode 100644 index 0000000000000000000000000000000000000000..eb7014a6b5480cc8123317379fcec9bda391c93f --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_high_school_geography.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _mg_template_yaml +process_docs: !function utils.process_high_school_geography +tag: global_mmlu_full_mg_social_sciences_tasks +task: global_mmlu_full_mg_high_school_geography diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_high_school_government_and_politics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_high_school_government_and_politics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..74c5fc18df2caa5da286ec066052a1b73906c4fc --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_high_school_government_and_politics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _mg_template_yaml +process_docs: !function utils.process_high_school_government_and_politics +tag: global_mmlu_full_mg_social_sciences_tasks +task: global_mmlu_full_mg_high_school_government_and_politics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_high_school_macroeconomics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_high_school_macroeconomics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..24631ff318e3f6f7004a9977cd4831c06188e183 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_high_school_macroeconomics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _mg_template_yaml +process_docs: !function utils.process_high_school_macroeconomics +tag: global_mmlu_full_mg_social_sciences_tasks +task: global_mmlu_full_mg_high_school_macroeconomics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_high_school_mathematics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_high_school_mathematics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..b9db4a0c2e13581833c88b96acb02bc02b3a7f65 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_high_school_mathematics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _mg_template_yaml +process_docs: !function utils.process_high_school_mathematics +tag: global_mmlu_full_mg_stem_tasks +task: global_mmlu_full_mg_high_school_mathematics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_high_school_microeconomics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_high_school_microeconomics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..f321b06ac63d32c971baec8863b9d91a91a73b08 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_high_school_microeconomics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _mg_template_yaml +process_docs: !function utils.process_high_school_microeconomics +tag: global_mmlu_full_mg_social_sciences_tasks +task: global_mmlu_full_mg_high_school_microeconomics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_high_school_physics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_high_school_physics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..bc25971e4c69aba3ddc63673b8044410570fc429 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_high_school_physics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _mg_template_yaml +process_docs: !function utils.process_high_school_physics +tag: global_mmlu_full_mg_stem_tasks +task: global_mmlu_full_mg_high_school_physics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_high_school_psychology.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_high_school_psychology.yaml new file mode 100644 index 0000000000000000000000000000000000000000..42cc39a8e78a32326c49315c3c9db5c25a23e820 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_high_school_psychology.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _mg_template_yaml +process_docs: !function utils.process_high_school_psychology +tag: global_mmlu_full_mg_social_sciences_tasks +task: global_mmlu_full_mg_high_school_psychology diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_high_school_statistics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_high_school_statistics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..08cf8671208f5a4d3b94e8056cea9d35cc0446c1 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_high_school_statistics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _mg_template_yaml +process_docs: !function utils.process_high_school_statistics +tag: global_mmlu_full_mg_stem_tasks +task: global_mmlu_full_mg_high_school_statistics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_high_school_us_history.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_high_school_us_history.yaml new file mode 100644 index 0000000000000000000000000000000000000000..87314a572bd152b27a37cf4e9f68c73a62f3060e --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_high_school_us_history.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _mg_template_yaml +process_docs: !function utils.process_high_school_us_history +tag: global_mmlu_full_mg_humanities_tasks +task: global_mmlu_full_mg_high_school_us_history diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_high_school_world_history.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_high_school_world_history.yaml new file mode 100644 index 0000000000000000000000000000000000000000..c341a243e9ca66ba06137a4934fc700c53993ebc --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_high_school_world_history.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _mg_template_yaml +process_docs: !function utils.process_high_school_world_history +tag: global_mmlu_full_mg_humanities_tasks +task: global_mmlu_full_mg_high_school_world_history diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_human_aging.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_human_aging.yaml new file mode 100644 index 0000000000000000000000000000000000000000..15375f9f0756d185c588cfd92abd58811df57dea --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_human_aging.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _mg_template_yaml +process_docs: !function utils.process_human_aging +tag: global_mmlu_full_mg_other_tasks +task: global_mmlu_full_mg_human_aging diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_human_sexuality.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_human_sexuality.yaml new file mode 100644 index 0000000000000000000000000000000000000000..21419b9b6dfcf1b8222d349a7acd254be0281a1b --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_human_sexuality.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _mg_template_yaml +process_docs: !function utils.process_human_sexuality +tag: global_mmlu_full_mg_social_sciences_tasks +task: global_mmlu_full_mg_human_sexuality diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_international_law.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_international_law.yaml new file mode 100644 index 0000000000000000000000000000000000000000..9d481339372484a7e9605dc623dd51b924b79cf1 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_international_law.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _mg_template_yaml +process_docs: !function utils.process_international_law +tag: global_mmlu_full_mg_humanities_tasks +task: global_mmlu_full_mg_international_law diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_jurisprudence.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_jurisprudence.yaml new file mode 100644 index 0000000000000000000000000000000000000000..f083a0ab591d6faf45d026043bf7ca452d1643bf --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_jurisprudence.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _mg_template_yaml +process_docs: !function utils.process_jurisprudence +tag: global_mmlu_full_mg_humanities_tasks +task: global_mmlu_full_mg_jurisprudence diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_logical_fallacies.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_logical_fallacies.yaml new file mode 100644 index 0000000000000000000000000000000000000000..57e2e731820760c2aba37c5b1bea43f8dcb80ad2 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_logical_fallacies.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _mg_template_yaml +process_docs: !function utils.process_logical_fallacies +tag: global_mmlu_full_mg_humanities_tasks +task: global_mmlu_full_mg_logical_fallacies diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_machine_learning.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_machine_learning.yaml new file mode 100644 index 0000000000000000000000000000000000000000..7609a09f687cdf4f1791cb7973e5658739f06d31 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_machine_learning.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _mg_template_yaml +process_docs: !function utils.process_machine_learning +tag: global_mmlu_full_mg_stem_tasks +task: global_mmlu_full_mg_machine_learning diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_management.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_management.yaml new file mode 100644 index 0000000000000000000000000000000000000000..becfe4b30db1ca6801bbb9a6c52f87d9c309d079 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_management.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _mg_template_yaml +process_docs: !function utils.process_management +tag: global_mmlu_full_mg_other_tasks +task: global_mmlu_full_mg_management diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_marketing.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_marketing.yaml new file mode 100644 index 0000000000000000000000000000000000000000..3765002b8210565c124d39a09a3477447506cc62 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_marketing.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _mg_template_yaml +process_docs: !function utils.process_marketing +tag: global_mmlu_full_mg_other_tasks +task: global_mmlu_full_mg_marketing diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_medical_genetics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_medical_genetics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..3f023ccd4d3d9ad1348b8ee1737c86ce167f05e5 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_medical_genetics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _mg_template_yaml +process_docs: !function utils.process_medical_genetics +tag: global_mmlu_full_mg_other_tasks +task: global_mmlu_full_mg_medical_genetics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_miscellaneous.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_miscellaneous.yaml new file mode 100644 index 0000000000000000000000000000000000000000..2993999d5a53f05e18e4be854c66eec9cbbcff1d --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_miscellaneous.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _mg_template_yaml +process_docs: !function utils.process_miscellaneous +tag: global_mmlu_full_mg_other_tasks +task: global_mmlu_full_mg_miscellaneous diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_moral_disputes.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_moral_disputes.yaml new file mode 100644 index 0000000000000000000000000000000000000000..fd430a0a171bbd9237ea960f8dc7b8c2e3c90b75 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_moral_disputes.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _mg_template_yaml +process_docs: !function utils.process_moral_disputes +tag: global_mmlu_full_mg_humanities_tasks +task: global_mmlu_full_mg_moral_disputes diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_moral_scenarios.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_moral_scenarios.yaml new file mode 100644 index 0000000000000000000000000000000000000000..c1b16e86512c0bc29969cc09f978d3888f4214d3 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_moral_scenarios.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _mg_template_yaml +process_docs: !function utils.process_moral_scenarios +tag: global_mmlu_full_mg_humanities_tasks +task: global_mmlu_full_mg_moral_scenarios diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_nutrition.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_nutrition.yaml new file mode 100644 index 0000000000000000000000000000000000000000..ab471f42cab9f636aa71bf8e238b9b4e8ac7623f --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_nutrition.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _mg_template_yaml +process_docs: !function utils.process_nutrition +tag: global_mmlu_full_mg_other_tasks +task: global_mmlu_full_mg_nutrition diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_philosophy.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_philosophy.yaml new file mode 100644 index 0000000000000000000000000000000000000000..f598830eda5f37fff7080609b667950753e971ec --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_philosophy.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _mg_template_yaml +process_docs: !function utils.process_philosophy +tag: global_mmlu_full_mg_humanities_tasks +task: global_mmlu_full_mg_philosophy diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_prehistory.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_prehistory.yaml new file mode 100644 index 0000000000000000000000000000000000000000..330f1f52aa90c715a3779255331265974ab6fac6 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_prehistory.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _mg_template_yaml +process_docs: !function utils.process_prehistory +tag: global_mmlu_full_mg_humanities_tasks +task: global_mmlu_full_mg_prehistory diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_professional_accounting.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_professional_accounting.yaml new file mode 100644 index 0000000000000000000000000000000000000000..694118d119b59f7270435d0a6ac0fbc258dca415 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_professional_accounting.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _mg_template_yaml +process_docs: !function utils.process_professional_accounting +tag: global_mmlu_full_mg_other_tasks +task: global_mmlu_full_mg_professional_accounting diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_professional_law.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_professional_law.yaml new file mode 100644 index 0000000000000000000000000000000000000000..fb6df92ac0f172469c43cda0fdbcb18cc7a2d650 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_professional_law.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _mg_template_yaml +process_docs: !function utils.process_professional_law +tag: global_mmlu_full_mg_humanities_tasks +task: global_mmlu_full_mg_professional_law diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_professional_medicine.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_professional_medicine.yaml new file mode 100644 index 0000000000000000000000000000000000000000..1de72b6b887ae1e9f081b13c7c00341d122cc2fc --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_professional_medicine.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _mg_template_yaml +process_docs: !function utils.process_professional_medicine +tag: global_mmlu_full_mg_other_tasks +task: global_mmlu_full_mg_professional_medicine diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_professional_psychology.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_professional_psychology.yaml new file mode 100644 index 0000000000000000000000000000000000000000..f922e1624b84e807143f7fc1bbb142710f338c93 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_professional_psychology.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _mg_template_yaml +process_docs: !function utils.process_professional_psychology +tag: global_mmlu_full_mg_social_sciences_tasks +task: global_mmlu_full_mg_professional_psychology diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_public_relations.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_public_relations.yaml new file mode 100644 index 0000000000000000000000000000000000000000..c829b89dc10281bb38ac778493bfc57a156dddb6 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_public_relations.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _mg_template_yaml +process_docs: !function utils.process_public_relations +tag: global_mmlu_full_mg_social_sciences_tasks +task: global_mmlu_full_mg_public_relations diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_security_studies.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_security_studies.yaml new file mode 100644 index 0000000000000000000000000000000000000000..362b4dbd13e1db423fb30a11b60720c10f37473f --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_security_studies.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _mg_template_yaml +process_docs: !function utils.process_security_studies +tag: global_mmlu_full_mg_social_sciences_tasks +task: global_mmlu_full_mg_security_studies diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_sociology.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_sociology.yaml new file mode 100644 index 0000000000000000000000000000000000000000..f0638cdb6d4535572c8a34a1c68ac25ea1c01e42 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_sociology.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _mg_template_yaml +process_docs: !function utils.process_sociology +tag: global_mmlu_full_mg_social_sciences_tasks +task: global_mmlu_full_mg_sociology diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_us_foreign_policy.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_us_foreign_policy.yaml new file mode 100644 index 0000000000000000000000000000000000000000..8ead541a433df13870d7f045b2564b4f0aa95f3f --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_us_foreign_policy.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _mg_template_yaml +process_docs: !function utils.process_us_foreign_policy +tag: global_mmlu_full_mg_social_sciences_tasks +task: global_mmlu_full_mg_us_foreign_policy diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_virology.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_virology.yaml new file mode 100644 index 0000000000000000000000000000000000000000..1ca09027f1cb3dc22cc1aa2bf8bebfa45a144276 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_virology.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _mg_template_yaml +process_docs: !function utils.process_virology +tag: global_mmlu_full_mg_other_tasks +task: global_mmlu_full_mg_virology diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_world_religions.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_world_religions.yaml new file mode 100644 index 0000000000000000000000000000000000000000..2bb64d70115db2e1f79637c8f0928b615cfca71c --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/global_mmlu_full_mg_world_religions.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _mg_template_yaml +process_docs: !function utils.process_world_religions +tag: global_mmlu_full_mg_humanities_tasks +task: global_mmlu_full_mg_world_religions diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/utils.py b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/utils.py new file mode 100644 index 0000000000000000000000000000000000000000..7df72cb061f0fecba46e15ff9b57552817979afb --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/mg/utils.py @@ -0,0 +1,73 @@ +from functools import partial + + +SUBJECTS = [ + "abstract_algebra", + "anatomy", + "astronomy", + "business_ethics", + "clinical_knowledge", + "college_biology", + "college_chemistry", + "college_computer_science", + "college_mathematics", + "college_medicine", + "college_physics", + "computer_security", + "conceptual_physics", + "econometrics", + "electrical_engineering", + "elementary_mathematics", + "formal_logic", + "global_facts", + "high_school_biology", + "high_school_chemistry", + "high_school_computer_science", + "high_school_european_history", + "high_school_geography", + "high_school_government_and_politics", + "high_school_macroeconomics", + "high_school_mathematics", + "high_school_microeconomics", + "high_school_physics", + "high_school_psychology", + "high_school_statistics", + "high_school_us_history", + "high_school_world_history", + "human_aging", + "human_sexuality", + "international_law", + "jurisprudence", + "logical_fallacies", + "machine_learning", + "management", + "marketing", + "medical_genetics", + "miscellaneous", + "moral_disputes", + "moral_scenarios", + "nutrition", + "philosophy", + "prehistory", + "professional_accounting", + "professional_law", + "professional_medicine", + "professional_psychology", + "public_relations", + "security_studies", + "sociology", + "us_foreign_policy", + "virology", + "world_religions", +] + + +def process_docs(dataset, subject): + return dataset.filter(lambda x: x["subject"] == subject) + + +process_functions = { + f"process_{subject}": partial(process_docs, subject=subject) for subject in SUBJECTS +} + +globals().update(process_functions) diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/_global_mmlu_full_ms.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/_global_mmlu_full_ms.yaml new file mode 100644 index 0000000000000000000000000000000000000000..a322dad5b8b2c05f3bc3b77ab3e1e58a31d1407e --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/_global_mmlu_full_ms.yaml @@ -0,0 +1,11 @@ +group: global_mmlu_full_ms +task: + - global_mmlu_full_ms_stem + - global_mmlu_full_ms_other + - global_mmlu_full_ms_social_sciences + - global_mmlu_full_ms_humanities +aggregate_metric_list: + - metric: acc + weight_by_size: True +metadata: + version: 0.0 diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/_global_mmlu_full_ms_humanities.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/_global_mmlu_full_ms_humanities.yaml new file mode 100644 index 0000000000000000000000000000000000000000..0641187b82aea23875e576e47e43ac55f0d7c2e8 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/_global_mmlu_full_ms_humanities.yaml @@ -0,0 +1,8 @@ +group: global_mmlu_full_ms_humanities +task: + - global_mmlu_full_ms_humanities_tasks +aggregate_metric_list: + - metric: acc + weight_by_size: true +metadata: + version: 0.0 diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/_global_mmlu_full_ms_other.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/_global_mmlu_full_ms_other.yaml new file mode 100644 index 0000000000000000000000000000000000000000..3d14420c01090b328a4f6117a62746e659868ec1 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/_global_mmlu_full_ms_other.yaml @@ -0,0 +1,8 @@ +group: global_mmlu_full_ms_other +task: + - global_mmlu_full_ms_other_tasks +aggregate_metric_list: + - metric: acc + weight_by_size: true +metadata: + version: 0.0 diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/_global_mmlu_full_ms_social_sciences.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/_global_mmlu_full_ms_social_sciences.yaml new file mode 100644 index 0000000000000000000000000000000000000000..3db339d79e64508a357a3de0dc62d9d5ae675a98 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/_global_mmlu_full_ms_social_sciences.yaml @@ -0,0 +1,8 @@ +group: global_mmlu_full_ms_social_sciences +task: + - global_mmlu_full_ms_social_sciences_tasks +aggregate_metric_list: + - metric: acc + weight_by_size: true +metadata: + version: 0.0 diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/_global_mmlu_full_ms_stem.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/_global_mmlu_full_ms_stem.yaml new file mode 100644 index 0000000000000000000000000000000000000000..68908e16ac2d25395f72b6a24da64a39eedb0f0c --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/_global_mmlu_full_ms_stem.yaml @@ -0,0 +1,8 @@ +group: global_mmlu_full_ms_stem +task: + - global_mmlu_full_ms_stem_tasks +aggregate_metric_list: + - metric: acc + weight_by_size: true +metadata: + version: 0.0 diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/_ms_template_yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/_ms_template_yaml new file mode 100644 index 0000000000000000000000000000000000000000..ba75026494233a4fde29e56c4de24b5d45669bc3 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/_ms_template_yaml @@ -0,0 +1,16 @@ +dataset_path: CohereForAI/Global-MMLU +dataset_name: ms +test_split: test +fewshot_split: dev +fewshot_config: + sampler: first_n +output_type: multiple_choice +doc_to_text: "{{question.strip()}}\nA. {{option_a}}\nB. {{option_b}}\nC. {{option_c}}\nD. {{option_d}}\nAnswer:" +doc_to_choice: ["A", "B", "C", "D"] +doc_to_target: answer +metric_list: + - metric: acc + aggregation: mean + higher_is_better: true +metadata: + version: 0.0 diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_abstract_algebra.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_abstract_algebra.yaml new file mode 100644 index 0000000000000000000000000000000000000000..ec791f2a08cca1bbcbe9d4f294e42aa93334a98f --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_abstract_algebra.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ms_template_yaml +process_docs: !function utils.process_abstract_algebra +tag: global_mmlu_full_ms_stem_tasks +task: global_mmlu_full_ms_abstract_algebra diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_anatomy.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_anatomy.yaml new file mode 100644 index 0000000000000000000000000000000000000000..35038beaff444dc1241589a79a1f1ed2b0f53557 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_anatomy.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ms_template_yaml +process_docs: !function utils.process_anatomy +tag: global_mmlu_full_ms_stem_tasks +task: global_mmlu_full_ms_anatomy diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_astronomy.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_astronomy.yaml new file mode 100644 index 0000000000000000000000000000000000000000..79fdcbdd0c9a02a1c11614a9785598c73a22b566 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_astronomy.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ms_template_yaml +process_docs: !function utils.process_astronomy +tag: global_mmlu_full_ms_stem_tasks +task: global_mmlu_full_ms_astronomy diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_business_ethics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_business_ethics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..ffd6195a9817669d3e1a76b4eaba3ec7af2de5bc --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_business_ethics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ms_template_yaml +process_docs: !function utils.process_business_ethics +tag: global_mmlu_full_ms_other_tasks +task: global_mmlu_full_ms_business_ethics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_clinical_knowledge.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_clinical_knowledge.yaml new file mode 100644 index 0000000000000000000000000000000000000000..4c69b82eb5f4b243e2b83d4257dcc0182a5d57bc --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_clinical_knowledge.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ms_template_yaml +process_docs: !function utils.process_clinical_knowledge +tag: global_mmlu_full_ms_other_tasks +task: global_mmlu_full_ms_clinical_knowledge diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_college_biology.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_college_biology.yaml new file mode 100644 index 0000000000000000000000000000000000000000..5821947946dc2f3bb6520331b5cb251a0fb46073 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_college_biology.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ms_template_yaml +process_docs: !function utils.process_college_biology +tag: global_mmlu_full_ms_stem_tasks +task: global_mmlu_full_ms_college_biology diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_college_chemistry.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_college_chemistry.yaml new file mode 100644 index 0000000000000000000000000000000000000000..35514b83405b51cb2870c012a443c2a11ed47983 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_college_chemistry.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ms_template_yaml +process_docs: !function utils.process_college_chemistry +tag: global_mmlu_full_ms_stem_tasks +task: global_mmlu_full_ms_college_chemistry diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_college_computer_science.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_college_computer_science.yaml new file mode 100644 index 0000000000000000000000000000000000000000..5e242b8bc260cabeedc1cf857377e26e935e441e --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_college_computer_science.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ms_template_yaml +process_docs: !function utils.process_college_computer_science +tag: global_mmlu_full_ms_stem_tasks +task: global_mmlu_full_ms_college_computer_science diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_college_mathematics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_college_mathematics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..07e107995767d4d7fd96ab4c53b08f40ca03d95d --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_college_mathematics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ms_template_yaml +process_docs: !function utils.process_college_mathematics +tag: global_mmlu_full_ms_stem_tasks +task: global_mmlu_full_ms_college_mathematics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_college_medicine.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_college_medicine.yaml new file mode 100644 index 0000000000000000000000000000000000000000..82822217f4782d36a8d10afdb9b08dd856a05978 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_college_medicine.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ms_template_yaml +process_docs: !function utils.process_college_medicine +tag: global_mmlu_full_ms_other_tasks +task: global_mmlu_full_ms_college_medicine diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_college_physics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_college_physics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..be20fa6cfec6f4c913270e96882861cdc93967d3 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_college_physics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ms_template_yaml +process_docs: !function utils.process_college_physics +tag: global_mmlu_full_ms_stem_tasks +task: global_mmlu_full_ms_college_physics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_computer_security.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_computer_security.yaml new file mode 100644 index 0000000000000000000000000000000000000000..2e886b501e59606923b487b90db6c8d74dc9b340 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_computer_security.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ms_template_yaml +process_docs: !function utils.process_computer_security +tag: global_mmlu_full_ms_stem_tasks +task: global_mmlu_full_ms_computer_security diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_conceptual_physics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_conceptual_physics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..2a2fb6da5343d528d5a2262cf4307a662e073dc1 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_conceptual_physics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ms_template_yaml +process_docs: !function utils.process_conceptual_physics +tag: global_mmlu_full_ms_stem_tasks +task: global_mmlu_full_ms_conceptual_physics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_econometrics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_econometrics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..efdffabf9e3ba98d2b110487cab97fab9c757cd0 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_econometrics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ms_template_yaml +process_docs: !function utils.process_econometrics +tag: global_mmlu_full_ms_social_sciences_tasks +task: global_mmlu_full_ms_econometrics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_electrical_engineering.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_electrical_engineering.yaml new file mode 100644 index 0000000000000000000000000000000000000000..80eba2e4795a06c026d2a54644eb5926dd9afcdd --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_electrical_engineering.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ms_template_yaml +process_docs: !function utils.process_electrical_engineering +tag: global_mmlu_full_ms_stem_tasks +task: global_mmlu_full_ms_electrical_engineering diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_elementary_mathematics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_elementary_mathematics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..1e6caf2688b9efd8012918b8a1959c0f2974b52b --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_elementary_mathematics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ms_template_yaml +process_docs: !function utils.process_elementary_mathematics +tag: global_mmlu_full_ms_stem_tasks +task: global_mmlu_full_ms_elementary_mathematics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_formal_logic.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_formal_logic.yaml new file mode 100644 index 0000000000000000000000000000000000000000..5914766232a2c7518e9517a3316d7fab569a752b --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_formal_logic.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ms_template_yaml +process_docs: !function utils.process_formal_logic +tag: global_mmlu_full_ms_humanities_tasks +task: global_mmlu_full_ms_formal_logic diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_global_facts.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_global_facts.yaml new file mode 100644 index 0000000000000000000000000000000000000000..6ac76cadf16c8d7feb8037ed42a50df02f7a2312 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_global_facts.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ms_template_yaml +process_docs: !function utils.process_global_facts +tag: global_mmlu_full_ms_other_tasks +task: global_mmlu_full_ms_global_facts diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_high_school_biology.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_high_school_biology.yaml new file mode 100644 index 0000000000000000000000000000000000000000..6be8ccfe1a25131ec4199bf16f735135ab566c52 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_high_school_biology.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ms_template_yaml +process_docs: !function utils.process_high_school_biology +tag: global_mmlu_full_ms_stem_tasks +task: global_mmlu_full_ms_high_school_biology diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_high_school_chemistry.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_high_school_chemistry.yaml new file mode 100644 index 0000000000000000000000000000000000000000..f01c29b671350403a193361f4cf6c90a717b6451 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_high_school_chemistry.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ms_template_yaml +process_docs: !function utils.process_high_school_chemistry +tag: global_mmlu_full_ms_stem_tasks +task: global_mmlu_full_ms_high_school_chemistry diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_high_school_computer_science.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_high_school_computer_science.yaml new file mode 100644 index 0000000000000000000000000000000000000000..b18e8cf85be18f2ac7980c628af29b55b1554fd1 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_high_school_computer_science.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ms_template_yaml +process_docs: !function utils.process_high_school_computer_science +tag: global_mmlu_full_ms_stem_tasks +task: global_mmlu_full_ms_high_school_computer_science diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_high_school_european_history.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_high_school_european_history.yaml new file mode 100644 index 0000000000000000000000000000000000000000..fdb41802d75ffd0c1ef875a31c04da74112ad7f4 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_high_school_european_history.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ms_template_yaml +process_docs: !function utils.process_high_school_european_history +tag: global_mmlu_full_ms_humanities_tasks +task: global_mmlu_full_ms_high_school_european_history diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_high_school_geography.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_high_school_geography.yaml new file mode 100644 index 0000000000000000000000000000000000000000..c4e44a60ef7a3d4f7ef7cae359c7f95934ab6286 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_high_school_geography.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ms_template_yaml +process_docs: !function utils.process_high_school_geography +tag: global_mmlu_full_ms_social_sciences_tasks +task: global_mmlu_full_ms_high_school_geography diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_high_school_government_and_politics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_high_school_government_and_politics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..0ebbfe6f4c20ed749b76b814b799926dddab4c46 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_high_school_government_and_politics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ms_template_yaml +process_docs: !function utils.process_high_school_government_and_politics +tag: global_mmlu_full_ms_social_sciences_tasks +task: global_mmlu_full_ms_high_school_government_and_politics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_high_school_macroeconomics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_high_school_macroeconomics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..f28f9a5ddbf9dde465b786c262aeb750f97dcb52 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_high_school_macroeconomics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ms_template_yaml +process_docs: !function utils.process_high_school_macroeconomics +tag: global_mmlu_full_ms_social_sciences_tasks +task: global_mmlu_full_ms_high_school_macroeconomics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_high_school_mathematics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_high_school_mathematics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..50a2552d0b7ee8f87cea9d4fceb9a8995505885d --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_high_school_mathematics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ms_template_yaml +process_docs: !function utils.process_high_school_mathematics +tag: global_mmlu_full_ms_stem_tasks +task: global_mmlu_full_ms_high_school_mathematics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_high_school_microeconomics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_high_school_microeconomics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..6747cd9da0599822287dcc572764f8ccddb2c38a --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_high_school_microeconomics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ms_template_yaml +process_docs: !function utils.process_high_school_microeconomics +tag: global_mmlu_full_ms_social_sciences_tasks +task: global_mmlu_full_ms_high_school_microeconomics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_high_school_physics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_high_school_physics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..aef3fee828e0a9f4944d9564df99a9d7281fb083 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_high_school_physics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ms_template_yaml +process_docs: !function utils.process_high_school_physics +tag: global_mmlu_full_ms_stem_tasks +task: global_mmlu_full_ms_high_school_physics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_high_school_psychology.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_high_school_psychology.yaml new file mode 100644 index 0000000000000000000000000000000000000000..3e8641e97d5b1d62c26a0b8d1afb0aaccf132a4e --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_high_school_psychology.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ms_template_yaml +process_docs: !function utils.process_high_school_psychology +tag: global_mmlu_full_ms_social_sciences_tasks +task: global_mmlu_full_ms_high_school_psychology diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_high_school_statistics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_high_school_statistics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..4aa7ba0057274449f3250892ba8e0b11a0ca711d --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_high_school_statistics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ms_template_yaml +process_docs: !function utils.process_high_school_statistics +tag: global_mmlu_full_ms_stem_tasks +task: global_mmlu_full_ms_high_school_statistics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_high_school_us_history.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_high_school_us_history.yaml new file mode 100644 index 0000000000000000000000000000000000000000..e6d1faabf2ff4fb989a09da310a472c66b0613fe --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_high_school_us_history.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ms_template_yaml +process_docs: !function utils.process_high_school_us_history +tag: global_mmlu_full_ms_humanities_tasks +task: global_mmlu_full_ms_high_school_us_history diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_high_school_world_history.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_high_school_world_history.yaml new file mode 100644 index 0000000000000000000000000000000000000000..4caf7e54727487e23fdcb6b7b5067ffe6cf59545 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_high_school_world_history.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ms_template_yaml +process_docs: !function utils.process_high_school_world_history +tag: global_mmlu_full_ms_humanities_tasks +task: global_mmlu_full_ms_high_school_world_history diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_human_aging.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_human_aging.yaml new file mode 100644 index 0000000000000000000000000000000000000000..5b2b5c5f4650a5ef51a387707bbaf3a40cfe7d01 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_human_aging.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ms_template_yaml +process_docs: !function utils.process_human_aging +tag: global_mmlu_full_ms_other_tasks +task: global_mmlu_full_ms_human_aging diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_human_sexuality.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_human_sexuality.yaml new file mode 100644 index 0000000000000000000000000000000000000000..2ddef17a710db537488125b00ae751a99cb3494a --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_human_sexuality.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ms_template_yaml +process_docs: !function utils.process_human_sexuality +tag: global_mmlu_full_ms_social_sciences_tasks +task: global_mmlu_full_ms_human_sexuality diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_international_law.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_international_law.yaml new file mode 100644 index 0000000000000000000000000000000000000000..61795f58976509acf5212696539476163145b7b7 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_international_law.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ms_template_yaml +process_docs: !function utils.process_international_law +tag: global_mmlu_full_ms_humanities_tasks +task: global_mmlu_full_ms_international_law diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_jurisprudence.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_jurisprudence.yaml new file mode 100644 index 0000000000000000000000000000000000000000..f2e96706b7cd686ca2048dd36c44e2978a90a4f6 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_jurisprudence.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ms_template_yaml +process_docs: !function utils.process_jurisprudence +tag: global_mmlu_full_ms_humanities_tasks +task: global_mmlu_full_ms_jurisprudence diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_logical_fallacies.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_logical_fallacies.yaml new file mode 100644 index 0000000000000000000000000000000000000000..1d142bde66593b7b106258a37534c3a678e7cf1d --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_logical_fallacies.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ms_template_yaml +process_docs: !function utils.process_logical_fallacies +tag: global_mmlu_full_ms_humanities_tasks +task: global_mmlu_full_ms_logical_fallacies diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_machine_learning.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_machine_learning.yaml new file mode 100644 index 0000000000000000000000000000000000000000..947240560206e5e186079c1f0e3f8a8c5c53193f --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_machine_learning.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ms_template_yaml +process_docs: !function utils.process_machine_learning +tag: global_mmlu_full_ms_stem_tasks +task: global_mmlu_full_ms_machine_learning diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_management.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_management.yaml new file mode 100644 index 0000000000000000000000000000000000000000..8ca04a13f3db9857da2dec99074052cccc5547f6 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_management.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ms_template_yaml +process_docs: !function utils.process_management +tag: global_mmlu_full_ms_other_tasks +task: global_mmlu_full_ms_management diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_marketing.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_marketing.yaml new file mode 100644 index 0000000000000000000000000000000000000000..ec0e44626cbb32df8bd06783583d5a88285268bb --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_marketing.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ms_template_yaml +process_docs: !function utils.process_marketing +tag: global_mmlu_full_ms_other_tasks +task: global_mmlu_full_ms_marketing diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_medical_genetics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_medical_genetics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..0f2b1eec7f0d0687e5bb730599968eed83864341 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_medical_genetics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ms_template_yaml +process_docs: !function utils.process_medical_genetics +tag: global_mmlu_full_ms_other_tasks +task: global_mmlu_full_ms_medical_genetics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_miscellaneous.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_miscellaneous.yaml new file mode 100644 index 0000000000000000000000000000000000000000..65da952e266b53e4c67a9c9d9881026337dd767a --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_miscellaneous.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ms_template_yaml +process_docs: !function utils.process_miscellaneous +tag: global_mmlu_full_ms_other_tasks +task: global_mmlu_full_ms_miscellaneous diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_moral_disputes.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_moral_disputes.yaml new file mode 100644 index 0000000000000000000000000000000000000000..399035f2c1567c14de08de33d331bb7cb62b03be --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_moral_disputes.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ms_template_yaml +process_docs: !function utils.process_moral_disputes +tag: global_mmlu_full_ms_humanities_tasks +task: global_mmlu_full_ms_moral_disputes diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_moral_scenarios.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_moral_scenarios.yaml new file mode 100644 index 0000000000000000000000000000000000000000..3bc74baa48b1b199a3cdfdf4ccff69fdd9246f3a --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_moral_scenarios.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ms_template_yaml +process_docs: !function utils.process_moral_scenarios +tag: global_mmlu_full_ms_humanities_tasks +task: global_mmlu_full_ms_moral_scenarios diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_nutrition.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_nutrition.yaml new file mode 100644 index 0000000000000000000000000000000000000000..300de6771aad1aae4994c08dc77ffeb78c3fc1d9 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_nutrition.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ms_template_yaml +process_docs: !function utils.process_nutrition +tag: global_mmlu_full_ms_other_tasks +task: global_mmlu_full_ms_nutrition diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_philosophy.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_philosophy.yaml new file mode 100644 index 0000000000000000000000000000000000000000..8f6eceaee97e08e1620a305fc0c2c8eb6cfcc8db --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_philosophy.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ms_template_yaml +process_docs: !function utils.process_philosophy +tag: global_mmlu_full_ms_humanities_tasks +task: global_mmlu_full_ms_philosophy diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_prehistory.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_prehistory.yaml new file mode 100644 index 0000000000000000000000000000000000000000..4c624fec5154688df3bfb017b79f48838e1c046f --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_prehistory.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ms_template_yaml +process_docs: !function utils.process_prehistory +tag: global_mmlu_full_ms_humanities_tasks +task: global_mmlu_full_ms_prehistory diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_professional_accounting.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_professional_accounting.yaml new file mode 100644 index 0000000000000000000000000000000000000000..9a06e7f922796015310ea788ccd805a4f7ebbafa --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_professional_accounting.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ms_template_yaml +process_docs: !function utils.process_professional_accounting +tag: global_mmlu_full_ms_other_tasks +task: global_mmlu_full_ms_professional_accounting diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_professional_law.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_professional_law.yaml new file mode 100644 index 0000000000000000000000000000000000000000..b3d5921a8ad8139a267cefa8d3b9850d0cbaceb2 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_professional_law.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ms_template_yaml +process_docs: !function utils.process_professional_law +tag: global_mmlu_full_ms_humanities_tasks +task: global_mmlu_full_ms_professional_law diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_professional_medicine.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_professional_medicine.yaml new file mode 100644 index 0000000000000000000000000000000000000000..0d9a58b7b8f81b127803dbb0eb81ad41dceb5d50 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_professional_medicine.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ms_template_yaml +process_docs: !function utils.process_professional_medicine +tag: global_mmlu_full_ms_other_tasks +task: global_mmlu_full_ms_professional_medicine diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_professional_psychology.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_professional_psychology.yaml new file mode 100644 index 0000000000000000000000000000000000000000..7f51baec9dbfbc1e5899f04f627d663ee1d88af4 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_professional_psychology.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ms_template_yaml +process_docs: !function utils.process_professional_psychology +tag: global_mmlu_full_ms_social_sciences_tasks +task: global_mmlu_full_ms_professional_psychology diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_public_relations.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_public_relations.yaml new file mode 100644 index 0000000000000000000000000000000000000000..c07cbdee8cb4a814d567d47b3cb20407ce95ba78 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_public_relations.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ms_template_yaml +process_docs: !function utils.process_public_relations +tag: global_mmlu_full_ms_social_sciences_tasks +task: global_mmlu_full_ms_public_relations diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_security_studies.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_security_studies.yaml new file mode 100644 index 0000000000000000000000000000000000000000..651cb72d4eb37961ebaca6a47ede956b727b596c --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_security_studies.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ms_template_yaml +process_docs: !function utils.process_security_studies +tag: global_mmlu_full_ms_social_sciences_tasks +task: global_mmlu_full_ms_security_studies diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_sociology.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_sociology.yaml new file mode 100644 index 0000000000000000000000000000000000000000..5aeb7efa4cd54a0267741924b853025065c8216e --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_sociology.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ms_template_yaml +process_docs: !function utils.process_sociology +tag: global_mmlu_full_ms_social_sciences_tasks +task: global_mmlu_full_ms_sociology diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_us_foreign_policy.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_us_foreign_policy.yaml new file mode 100644 index 0000000000000000000000000000000000000000..ecbf57059919c16149449d97829f8a281acb5d61 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_us_foreign_policy.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ms_template_yaml +process_docs: !function utils.process_us_foreign_policy +tag: global_mmlu_full_ms_social_sciences_tasks +task: global_mmlu_full_ms_us_foreign_policy diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_virology.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_virology.yaml new file mode 100644 index 0000000000000000000000000000000000000000..fbdd5e256571de2874c574db059771d6ecc40da6 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_virology.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ms_template_yaml +process_docs: !function utils.process_virology +tag: global_mmlu_full_ms_other_tasks +task: global_mmlu_full_ms_virology diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_world_religions.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_world_religions.yaml new file mode 100644 index 0000000000000000000000000000000000000000..32b350294c8410005952d248c61d9a84375d8c21 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/global_mmlu_full_ms_world_religions.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ms_template_yaml +process_docs: !function utils.process_world_religions +tag: global_mmlu_full_ms_humanities_tasks +task: global_mmlu_full_ms_world_religions diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/utils.py b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/utils.py new file mode 100644 index 0000000000000000000000000000000000000000..7df72cb061f0fecba46e15ff9b57552817979afb --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ms/utils.py @@ -0,0 +1,73 @@ +from functools import partial + + +SUBJECTS = [ + "abstract_algebra", + "anatomy", + "astronomy", + "business_ethics", + "clinical_knowledge", + "college_biology", + "college_chemistry", + "college_computer_science", + "college_mathematics", + "college_medicine", + "college_physics", + "computer_security", + "conceptual_physics", + "econometrics", + "electrical_engineering", + "elementary_mathematics", + "formal_logic", + "global_facts", + "high_school_biology", + "high_school_chemistry", + "high_school_computer_science", + "high_school_european_history", + "high_school_geography", + "high_school_government_and_politics", + "high_school_macroeconomics", + "high_school_mathematics", + "high_school_microeconomics", + "high_school_physics", + "high_school_psychology", + "high_school_statistics", + "high_school_us_history", + "high_school_world_history", + "human_aging", + "human_sexuality", + "international_law", + "jurisprudence", + "logical_fallacies", + "machine_learning", + "management", + "marketing", + "medical_genetics", + "miscellaneous", + "moral_disputes", + "moral_scenarios", + "nutrition", + "philosophy", + "prehistory", + "professional_accounting", + "professional_law", + "professional_medicine", + "professional_psychology", + "public_relations", + "security_studies", + "sociology", + "us_foreign_policy", + "virology", + "world_religions", +] + + +def process_docs(dataset, subject): + return dataset.filter(lambda x: x["subject"] == subject) + + +process_functions = { + f"process_{subject}": partial(process_docs, subject=subject) for subject in SUBJECTS +} + +globals().update(process_functions) diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/_global_mmlu_full_ne.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/_global_mmlu_full_ne.yaml new file mode 100644 index 0000000000000000000000000000000000000000..7e9c77ce6c1098ac620f41e5f49ae8b1bfe36b03 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/_global_mmlu_full_ne.yaml @@ -0,0 +1,11 @@ +group: global_mmlu_full_ne +task: + - global_mmlu_full_ne_stem + - global_mmlu_full_ne_other + - global_mmlu_full_ne_social_sciences + - global_mmlu_full_ne_humanities +aggregate_metric_list: + - metric: acc + weight_by_size: True +metadata: + version: 0.0 diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/_global_mmlu_full_ne_humanities.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/_global_mmlu_full_ne_humanities.yaml new file mode 100644 index 0000000000000000000000000000000000000000..fef749dbfa7cd9fc848e010ef5c318b226b6f194 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/_global_mmlu_full_ne_humanities.yaml @@ -0,0 +1,8 @@ +group: global_mmlu_full_ne_humanities +task: + - global_mmlu_full_ne_humanities_tasks +aggregate_metric_list: + - metric: acc + weight_by_size: true +metadata: + version: 0.0 diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/_global_mmlu_full_ne_other.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/_global_mmlu_full_ne_other.yaml new file mode 100644 index 0000000000000000000000000000000000000000..0d3dfbd466c3b8aefdbb14d324d0de1b3653cb34 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/_global_mmlu_full_ne_other.yaml @@ -0,0 +1,8 @@ +group: global_mmlu_full_ne_other +task: + - global_mmlu_full_ne_other_tasks +aggregate_metric_list: + - metric: acc + weight_by_size: true +metadata: + version: 0.0 diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/_global_mmlu_full_ne_social_sciences.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/_global_mmlu_full_ne_social_sciences.yaml new file mode 100644 index 0000000000000000000000000000000000000000..f1f09f00211510df0ac89d1f8cf28baa0eedf214 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/_global_mmlu_full_ne_social_sciences.yaml @@ -0,0 +1,8 @@ +group: global_mmlu_full_ne_social_sciences +task: + - global_mmlu_full_ne_social_sciences_tasks +aggregate_metric_list: + - metric: acc + weight_by_size: true +metadata: + version: 0.0 diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/_global_mmlu_full_ne_stem.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/_global_mmlu_full_ne_stem.yaml new file mode 100644 index 0000000000000000000000000000000000000000..eebc1cac3767ca67a10edb6c38fd70b5c54dc15a --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/_global_mmlu_full_ne_stem.yaml @@ -0,0 +1,8 @@ +group: global_mmlu_full_ne_stem +task: + - global_mmlu_full_ne_stem_tasks +aggregate_metric_list: + - metric: acc + weight_by_size: true +metadata: + version: 0.0 diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/_ne_template_yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/_ne_template_yaml new file mode 100644 index 0000000000000000000000000000000000000000..25f8daecd02bbb0414f665fce6ccda58f1cb0baf --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/_ne_template_yaml @@ -0,0 +1,16 @@ +dataset_path: CohereForAI/Global-MMLU +dataset_name: ne +test_split: test +fewshot_split: dev +fewshot_config: + sampler: first_n +output_type: multiple_choice +doc_to_text: "{{question.strip()}}\nA. {{option_a}}\nB. {{option_b}}\nC. {{option_c}}\nD. {{option_d}}\nAnswer:" +doc_to_choice: ["A", "B", "C", "D"] +doc_to_target: answer +metric_list: + - metric: acc + aggregation: mean + higher_is_better: true +metadata: + version: 0.0 diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_abstract_algebra.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_abstract_algebra.yaml new file mode 100644 index 0000000000000000000000000000000000000000..48bf7bb188c221ebffcb2a729e5fc86649b7f05b --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_abstract_algebra.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ne_template_yaml +process_docs: !function utils.process_abstract_algebra +tag: global_mmlu_full_ne_stem_tasks +task: global_mmlu_full_ne_abstract_algebra diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_anatomy.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_anatomy.yaml new file mode 100644 index 0000000000000000000000000000000000000000..0f66f8ec7803650ac5cff6a60dfc9bb9f8388fbe --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_anatomy.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ne_template_yaml +process_docs: !function utils.process_anatomy +tag: global_mmlu_full_ne_stem_tasks +task: global_mmlu_full_ne_anatomy diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_astronomy.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_astronomy.yaml new file mode 100644 index 0000000000000000000000000000000000000000..a02aaf30cd78e9cbb9d24aaf5f3787f8a6749a21 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_astronomy.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ne_template_yaml +process_docs: !function utils.process_astronomy +tag: global_mmlu_full_ne_stem_tasks +task: global_mmlu_full_ne_astronomy diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_business_ethics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_business_ethics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..d87f5b985c6e8a7f91b0b6646a72d290e72a1112 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_business_ethics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ne_template_yaml +process_docs: !function utils.process_business_ethics +tag: global_mmlu_full_ne_other_tasks +task: global_mmlu_full_ne_business_ethics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_clinical_knowledge.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_clinical_knowledge.yaml new file mode 100644 index 0000000000000000000000000000000000000000..f27eb4e2d358ba86a0d7567f2affc416143989bd --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_clinical_knowledge.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ne_template_yaml +process_docs: !function utils.process_clinical_knowledge +tag: global_mmlu_full_ne_other_tasks +task: global_mmlu_full_ne_clinical_knowledge diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_college_biology.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_college_biology.yaml new file mode 100644 index 0000000000000000000000000000000000000000..d26edef8af4d7dc876a808a278ef2fe63d4ed0ea --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_college_biology.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ne_template_yaml +process_docs: !function utils.process_college_biology +tag: global_mmlu_full_ne_stem_tasks +task: global_mmlu_full_ne_college_biology diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_college_chemistry.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_college_chemistry.yaml new file mode 100644 index 0000000000000000000000000000000000000000..88b8bd86f4f4345ca2a250b3f12def25dd018aaf --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_college_chemistry.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ne_template_yaml +process_docs: !function utils.process_college_chemistry +tag: global_mmlu_full_ne_stem_tasks +task: global_mmlu_full_ne_college_chemistry diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_college_computer_science.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_college_computer_science.yaml new file mode 100644 index 0000000000000000000000000000000000000000..51909ffc840297e9f9028e0086fc159eea82181e --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_college_computer_science.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ne_template_yaml +process_docs: !function utils.process_college_computer_science +tag: global_mmlu_full_ne_stem_tasks +task: global_mmlu_full_ne_college_computer_science diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_college_mathematics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_college_mathematics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..40b9cb7963041a34533068f20110e20b575049cb --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_college_mathematics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ne_template_yaml +process_docs: !function utils.process_college_mathematics +tag: global_mmlu_full_ne_stem_tasks +task: global_mmlu_full_ne_college_mathematics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_college_medicine.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_college_medicine.yaml new file mode 100644 index 0000000000000000000000000000000000000000..81f81f84437cc9c0dcf847b08d2969a9099554da --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_college_medicine.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ne_template_yaml +process_docs: !function utils.process_college_medicine +tag: global_mmlu_full_ne_other_tasks +task: global_mmlu_full_ne_college_medicine diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_college_physics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_college_physics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..09798c096fa17206a0a96b9765ce2b4d05a74684 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_college_physics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ne_template_yaml +process_docs: !function utils.process_college_physics +tag: global_mmlu_full_ne_stem_tasks +task: global_mmlu_full_ne_college_physics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_computer_security.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_computer_security.yaml new file mode 100644 index 0000000000000000000000000000000000000000..49d89dd43577cf8931b0209d826e07fbda631c0e --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_computer_security.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ne_template_yaml +process_docs: !function utils.process_computer_security +tag: global_mmlu_full_ne_stem_tasks +task: global_mmlu_full_ne_computer_security diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_conceptual_physics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_conceptual_physics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..94bfec4a90dfffb4a44a71cc28061587bb1eb2ce --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_conceptual_physics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ne_template_yaml +process_docs: !function utils.process_conceptual_physics +tag: global_mmlu_full_ne_stem_tasks +task: global_mmlu_full_ne_conceptual_physics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_econometrics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_econometrics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..81d6ed98e045976c34e0043e4b3c3026e958d695 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_econometrics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ne_template_yaml +process_docs: !function utils.process_econometrics +tag: global_mmlu_full_ne_social_sciences_tasks +task: global_mmlu_full_ne_econometrics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_electrical_engineering.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_electrical_engineering.yaml new file mode 100644 index 0000000000000000000000000000000000000000..73ad1a34a39b3944ef2634c780527caa3a82d471 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_electrical_engineering.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ne_template_yaml +process_docs: !function utils.process_electrical_engineering +tag: global_mmlu_full_ne_stem_tasks +task: global_mmlu_full_ne_electrical_engineering diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_elementary_mathematics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_elementary_mathematics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..cbc3bacd0839d6c0fbb8b86eb7501c1f92b303e8 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_elementary_mathematics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ne_template_yaml +process_docs: !function utils.process_elementary_mathematics +tag: global_mmlu_full_ne_stem_tasks +task: global_mmlu_full_ne_elementary_mathematics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_formal_logic.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_formal_logic.yaml new file mode 100644 index 0000000000000000000000000000000000000000..225da2fb5bd475a35868879e26670145d6e0b880 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_formal_logic.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ne_template_yaml +process_docs: !function utils.process_formal_logic +tag: global_mmlu_full_ne_humanities_tasks +task: global_mmlu_full_ne_formal_logic diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_global_facts.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_global_facts.yaml new file mode 100644 index 0000000000000000000000000000000000000000..6f5e9f1b4e447bb7658d73e7f48e19b16d55ff7f --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_global_facts.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ne_template_yaml +process_docs: !function utils.process_global_facts +tag: global_mmlu_full_ne_other_tasks +task: global_mmlu_full_ne_global_facts diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_high_school_biology.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_high_school_biology.yaml new file mode 100644 index 0000000000000000000000000000000000000000..a8c0436a44f598b83d8935e28ef61e8708415cdf --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_high_school_biology.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ne_template_yaml +process_docs: !function utils.process_high_school_biology +tag: global_mmlu_full_ne_stem_tasks +task: global_mmlu_full_ne_high_school_biology diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_high_school_chemistry.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_high_school_chemistry.yaml new file mode 100644 index 0000000000000000000000000000000000000000..405661c7e2af99823f8352f098a24755c610e9a2 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_high_school_chemistry.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ne_template_yaml +process_docs: !function utils.process_high_school_chemistry +tag: global_mmlu_full_ne_stem_tasks +task: global_mmlu_full_ne_high_school_chemistry diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_high_school_computer_science.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_high_school_computer_science.yaml new file mode 100644 index 0000000000000000000000000000000000000000..6cff5ba64b352c0d7f391492e95d173e641144f5 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_high_school_computer_science.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ne_template_yaml +process_docs: !function utils.process_high_school_computer_science +tag: global_mmlu_full_ne_stem_tasks +task: global_mmlu_full_ne_high_school_computer_science diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_high_school_european_history.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_high_school_european_history.yaml new file mode 100644 index 0000000000000000000000000000000000000000..4f7eb3ee33e2449dcafee69e07263f395fdd72b0 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_high_school_european_history.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ne_template_yaml +process_docs: !function utils.process_high_school_european_history +tag: global_mmlu_full_ne_humanities_tasks +task: global_mmlu_full_ne_high_school_european_history diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_high_school_geography.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_high_school_geography.yaml new file mode 100644 index 0000000000000000000000000000000000000000..0453e51d9d668942973887dc64ea15d0d93584a1 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_high_school_geography.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ne_template_yaml +process_docs: !function utils.process_high_school_geography +tag: global_mmlu_full_ne_social_sciences_tasks +task: global_mmlu_full_ne_high_school_geography diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_high_school_government_and_politics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_high_school_government_and_politics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..05710100729043b043e6b6c4c8e87e418b168e36 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_high_school_government_and_politics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ne_template_yaml +process_docs: !function utils.process_high_school_government_and_politics +tag: global_mmlu_full_ne_social_sciences_tasks +task: global_mmlu_full_ne_high_school_government_and_politics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_high_school_macroeconomics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_high_school_macroeconomics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..fd68d5f041e914b82b4cfc5df2b2ce0f21baa2a8 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_high_school_macroeconomics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ne_template_yaml +process_docs: !function utils.process_high_school_macroeconomics +tag: global_mmlu_full_ne_social_sciences_tasks +task: global_mmlu_full_ne_high_school_macroeconomics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_high_school_mathematics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_high_school_mathematics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..39ef0a58f7fd865822aa293ec6becd3e74559809 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_high_school_mathematics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ne_template_yaml +process_docs: !function utils.process_high_school_mathematics +tag: global_mmlu_full_ne_stem_tasks +task: global_mmlu_full_ne_high_school_mathematics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_high_school_microeconomics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_high_school_microeconomics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..535a391821a13f937d59174163b0a481758969ae --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_high_school_microeconomics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ne_template_yaml +process_docs: !function utils.process_high_school_microeconomics +tag: global_mmlu_full_ne_social_sciences_tasks +task: global_mmlu_full_ne_high_school_microeconomics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_high_school_physics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_high_school_physics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..f355dad253203b92e2f7903d12c8353178ab3232 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_high_school_physics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ne_template_yaml +process_docs: !function utils.process_high_school_physics +tag: global_mmlu_full_ne_stem_tasks +task: global_mmlu_full_ne_high_school_physics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_high_school_psychology.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_high_school_psychology.yaml new file mode 100644 index 0000000000000000000000000000000000000000..a52d7a017bf063a16cde6b3a65866785a65c4dbf --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_high_school_psychology.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ne_template_yaml +process_docs: !function utils.process_high_school_psychology +tag: global_mmlu_full_ne_social_sciences_tasks +task: global_mmlu_full_ne_high_school_psychology diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_high_school_statistics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_high_school_statistics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..5a2564200d4cdc0a0b0c5dd68ab4421861f1c06c --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_high_school_statistics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ne_template_yaml +process_docs: !function utils.process_high_school_statistics +tag: global_mmlu_full_ne_stem_tasks +task: global_mmlu_full_ne_high_school_statistics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_high_school_us_history.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_high_school_us_history.yaml new file mode 100644 index 0000000000000000000000000000000000000000..9e1199b17fb77ff7bf7c7eef871c8fa2d43fe5ea --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_high_school_us_history.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ne_template_yaml +process_docs: !function utils.process_high_school_us_history +tag: global_mmlu_full_ne_humanities_tasks +task: global_mmlu_full_ne_high_school_us_history diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_high_school_world_history.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_high_school_world_history.yaml new file mode 100644 index 0000000000000000000000000000000000000000..afc2135b37ddae395d6d4f0864f2ce42e23a535b --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_high_school_world_history.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ne_template_yaml +process_docs: !function utils.process_high_school_world_history +tag: global_mmlu_full_ne_humanities_tasks +task: global_mmlu_full_ne_high_school_world_history diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_human_aging.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_human_aging.yaml new file mode 100644 index 0000000000000000000000000000000000000000..18450534654c637b06c56836a9eaaa44662718ba --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_human_aging.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ne_template_yaml +process_docs: !function utils.process_human_aging +tag: global_mmlu_full_ne_other_tasks +task: global_mmlu_full_ne_human_aging diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_human_sexuality.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_human_sexuality.yaml new file mode 100644 index 0000000000000000000000000000000000000000..7d23b8397bd04a04b2a3d50f750b39996cfdbbcb --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_human_sexuality.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ne_template_yaml +process_docs: !function utils.process_human_sexuality +tag: global_mmlu_full_ne_social_sciences_tasks +task: global_mmlu_full_ne_human_sexuality diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_international_law.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_international_law.yaml new file mode 100644 index 0000000000000000000000000000000000000000..5be599d20c20a2b87ee2043abe7df0d07eca7326 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_international_law.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ne_template_yaml +process_docs: !function utils.process_international_law +tag: global_mmlu_full_ne_humanities_tasks +task: global_mmlu_full_ne_international_law diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_jurisprudence.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_jurisprudence.yaml new file mode 100644 index 0000000000000000000000000000000000000000..180a397c62f976822662b1b3997b352b8bf09d23 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_jurisprudence.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ne_template_yaml +process_docs: !function utils.process_jurisprudence +tag: global_mmlu_full_ne_humanities_tasks +task: global_mmlu_full_ne_jurisprudence diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_logical_fallacies.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_logical_fallacies.yaml new file mode 100644 index 0000000000000000000000000000000000000000..3aa369a96354c77c7e6ee0fb2b9e57f99887d8b4 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_logical_fallacies.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ne_template_yaml +process_docs: !function utils.process_logical_fallacies +tag: global_mmlu_full_ne_humanities_tasks +task: global_mmlu_full_ne_logical_fallacies diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_machine_learning.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_machine_learning.yaml new file mode 100644 index 0000000000000000000000000000000000000000..4e08abdac804b60935363639309106374cdaaef4 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_machine_learning.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ne_template_yaml +process_docs: !function utils.process_machine_learning +tag: global_mmlu_full_ne_stem_tasks +task: global_mmlu_full_ne_machine_learning diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_management.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_management.yaml new file mode 100644 index 0000000000000000000000000000000000000000..e44c5be63b6bec1011631761b40c32f93c4b775e --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_management.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ne_template_yaml +process_docs: !function utils.process_management +tag: global_mmlu_full_ne_other_tasks +task: global_mmlu_full_ne_management diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_marketing.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_marketing.yaml new file mode 100644 index 0000000000000000000000000000000000000000..10f7daa274c1b821b33b73878f22a186c5629715 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_marketing.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ne_template_yaml +process_docs: !function utils.process_marketing +tag: global_mmlu_full_ne_other_tasks +task: global_mmlu_full_ne_marketing diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_medical_genetics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_medical_genetics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..8139b1f7a65c5ff5668d79e8d3f0b19cf1d866e9 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_medical_genetics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ne_template_yaml +process_docs: !function utils.process_medical_genetics +tag: global_mmlu_full_ne_other_tasks +task: global_mmlu_full_ne_medical_genetics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_miscellaneous.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_miscellaneous.yaml new file mode 100644 index 0000000000000000000000000000000000000000..cb1bf905061a29950ba1ed2b50678fc8b25cc489 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_miscellaneous.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ne_template_yaml +process_docs: !function utils.process_miscellaneous +tag: global_mmlu_full_ne_other_tasks +task: global_mmlu_full_ne_miscellaneous diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_moral_disputes.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_moral_disputes.yaml new file mode 100644 index 0000000000000000000000000000000000000000..1b74fb365022fb4eda448dc1fc77258780c49428 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_moral_disputes.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ne_template_yaml +process_docs: !function utils.process_moral_disputes +tag: global_mmlu_full_ne_humanities_tasks +task: global_mmlu_full_ne_moral_disputes diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_moral_scenarios.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_moral_scenarios.yaml new file mode 100644 index 0000000000000000000000000000000000000000..91f8f06c5a769c701fdf7f7bbfd0955d9b837a05 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_moral_scenarios.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ne_template_yaml +process_docs: !function utils.process_moral_scenarios +tag: global_mmlu_full_ne_humanities_tasks +task: global_mmlu_full_ne_moral_scenarios diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_nutrition.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_nutrition.yaml new file mode 100644 index 0000000000000000000000000000000000000000..575f0e45506432d31b4c34c5d602b449a1e3d8f0 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_nutrition.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ne_template_yaml +process_docs: !function utils.process_nutrition +tag: global_mmlu_full_ne_other_tasks +task: global_mmlu_full_ne_nutrition diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_philosophy.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_philosophy.yaml new file mode 100644 index 0000000000000000000000000000000000000000..95fdd0eb343fc2a738c900782c9ce37dcb8c429a --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_philosophy.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ne_template_yaml +process_docs: !function utils.process_philosophy +tag: global_mmlu_full_ne_humanities_tasks +task: global_mmlu_full_ne_philosophy diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_prehistory.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_prehistory.yaml new file mode 100644 index 0000000000000000000000000000000000000000..e6e5c706c0b37d1d0a176f0d43e6fc41c8a47485 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_prehistory.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ne_template_yaml +process_docs: !function utils.process_prehistory +tag: global_mmlu_full_ne_humanities_tasks +task: global_mmlu_full_ne_prehistory diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_professional_accounting.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_professional_accounting.yaml new file mode 100644 index 0000000000000000000000000000000000000000..718cedee02328945a0e05bcb590ab2ba158af9f5 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_professional_accounting.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ne_template_yaml +process_docs: !function utils.process_professional_accounting +tag: global_mmlu_full_ne_other_tasks +task: global_mmlu_full_ne_professional_accounting diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_professional_law.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_professional_law.yaml new file mode 100644 index 0000000000000000000000000000000000000000..89c70160a6bbbba493f5e3c702ea02ebd850f94e --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_professional_law.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ne_template_yaml +process_docs: !function utils.process_professional_law +tag: global_mmlu_full_ne_humanities_tasks +task: global_mmlu_full_ne_professional_law diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_professional_medicine.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_professional_medicine.yaml new file mode 100644 index 0000000000000000000000000000000000000000..a366e0c43a3ccadd06b6e645d3f11da62cb67486 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_professional_medicine.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ne_template_yaml +process_docs: !function utils.process_professional_medicine +tag: global_mmlu_full_ne_other_tasks +task: global_mmlu_full_ne_professional_medicine diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_professional_psychology.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_professional_psychology.yaml new file mode 100644 index 0000000000000000000000000000000000000000..649e53435a1e0842ed0280efd91ed9ed8c32c822 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_professional_psychology.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ne_template_yaml +process_docs: !function utils.process_professional_psychology +tag: global_mmlu_full_ne_social_sciences_tasks +task: global_mmlu_full_ne_professional_psychology diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_public_relations.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_public_relations.yaml new file mode 100644 index 0000000000000000000000000000000000000000..37f2ddeab876d0c1861475dafb53c52cb20e673c --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_public_relations.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ne_template_yaml +process_docs: !function utils.process_public_relations +tag: global_mmlu_full_ne_social_sciences_tasks +task: global_mmlu_full_ne_public_relations diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_security_studies.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_security_studies.yaml new file mode 100644 index 0000000000000000000000000000000000000000..55f809041b2b726de52d63ebecec0a584f361317 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_security_studies.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ne_template_yaml +process_docs: !function utils.process_security_studies +tag: global_mmlu_full_ne_social_sciences_tasks +task: global_mmlu_full_ne_security_studies diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_sociology.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_sociology.yaml new file mode 100644 index 0000000000000000000000000000000000000000..78161d5a113e867abf882a1824a6a6a8fa590d4b --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_sociology.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ne_template_yaml +process_docs: !function utils.process_sociology +tag: global_mmlu_full_ne_social_sciences_tasks +task: global_mmlu_full_ne_sociology diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_us_foreign_policy.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_us_foreign_policy.yaml new file mode 100644 index 0000000000000000000000000000000000000000..c38f59c405afe4b39415a5ff16d6f4083a23be1a --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_us_foreign_policy.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ne_template_yaml +process_docs: !function utils.process_us_foreign_policy +tag: global_mmlu_full_ne_social_sciences_tasks +task: global_mmlu_full_ne_us_foreign_policy diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_virology.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_virology.yaml new file mode 100644 index 0000000000000000000000000000000000000000..0c15808fdc84a2c3b216833352c94731e83cb64d --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_virology.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ne_template_yaml +process_docs: !function utils.process_virology +tag: global_mmlu_full_ne_other_tasks +task: global_mmlu_full_ne_virology diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_world_religions.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_world_religions.yaml new file mode 100644 index 0000000000000000000000000000000000000000..5c6163f12c9d3d90ffb4b9576eac1f809bb37e03 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/global_mmlu_full_ne_world_religions.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ne_template_yaml +process_docs: !function utils.process_world_religions +tag: global_mmlu_full_ne_humanities_tasks +task: global_mmlu_full_ne_world_religions diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/utils.py b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/utils.py new file mode 100644 index 0000000000000000000000000000000000000000..7df72cb061f0fecba46e15ff9b57552817979afb --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ne/utils.py @@ -0,0 +1,73 @@ +from functools import partial + + +SUBJECTS = [ + "abstract_algebra", + "anatomy", + "astronomy", + "business_ethics", + "clinical_knowledge", + "college_biology", + "college_chemistry", + "college_computer_science", + "college_mathematics", + "college_medicine", + "college_physics", + "computer_security", + "conceptual_physics", + "econometrics", + "electrical_engineering", + "elementary_mathematics", + "formal_logic", + "global_facts", + "high_school_biology", + "high_school_chemistry", + "high_school_computer_science", + "high_school_european_history", + "high_school_geography", + "high_school_government_and_politics", + "high_school_macroeconomics", + "high_school_mathematics", + "high_school_microeconomics", + "high_school_physics", + "high_school_psychology", + "high_school_statistics", + "high_school_us_history", + "high_school_world_history", + "human_aging", + "human_sexuality", + "international_law", + "jurisprudence", + "logical_fallacies", + "machine_learning", + "management", + "marketing", + "medical_genetics", + "miscellaneous", + "moral_disputes", + "moral_scenarios", + "nutrition", + "philosophy", + "prehistory", + "professional_accounting", + "professional_law", + "professional_medicine", + "professional_psychology", + "public_relations", + "security_studies", + "sociology", + "us_foreign_policy", + "virology", + "world_religions", +] + + +def process_docs(dataset, subject): + return dataset.filter(lambda x: x["subject"] == subject) + + +process_functions = { + f"process_{subject}": partial(process_docs, subject=subject) for subject in SUBJECTS +} + +globals().update(process_functions) diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/_global_mmlu_full_nl.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/_global_mmlu_full_nl.yaml new file mode 100644 index 0000000000000000000000000000000000000000..c293a558a102282b53df954335faf7bb474f69b7 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/_global_mmlu_full_nl.yaml @@ -0,0 +1,11 @@ +group: global_mmlu_full_nl +task: + - global_mmlu_full_nl_stem + - global_mmlu_full_nl_other + - global_mmlu_full_nl_social_sciences + - global_mmlu_full_nl_humanities +aggregate_metric_list: + - metric: acc + weight_by_size: True +metadata: + version: 0.0 diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/_global_mmlu_full_nl_humanities.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/_global_mmlu_full_nl_humanities.yaml new file mode 100644 index 0000000000000000000000000000000000000000..656a421bfb7dc29eebfe0013b7c69b331fe04c6d --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/_global_mmlu_full_nl_humanities.yaml @@ -0,0 +1,8 @@ +group: global_mmlu_full_nl_humanities +task: + - global_mmlu_full_nl_humanities_tasks +aggregate_metric_list: + - metric: acc + weight_by_size: true +metadata: + version: 0.0 diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/_global_mmlu_full_nl_other.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/_global_mmlu_full_nl_other.yaml new file mode 100644 index 0000000000000000000000000000000000000000..23a42201ec4f8212e836727b434804977111cf7a --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/_global_mmlu_full_nl_other.yaml @@ -0,0 +1,8 @@ +group: global_mmlu_full_nl_other +task: + - global_mmlu_full_nl_other_tasks +aggregate_metric_list: + - metric: acc + weight_by_size: true +metadata: + version: 0.0 diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/_global_mmlu_full_nl_social_sciences.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/_global_mmlu_full_nl_social_sciences.yaml new file mode 100644 index 0000000000000000000000000000000000000000..afba5678d182bfcbcef3bf227a1800dfda12b535 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/_global_mmlu_full_nl_social_sciences.yaml @@ -0,0 +1,8 @@ +group: global_mmlu_full_nl_social_sciences +task: + - global_mmlu_full_nl_social_sciences_tasks +aggregate_metric_list: + - metric: acc + weight_by_size: true +metadata: + version: 0.0 diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/_global_mmlu_full_nl_stem.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/_global_mmlu_full_nl_stem.yaml new file mode 100644 index 0000000000000000000000000000000000000000..9658b13ed512182d510fd72c60cf1eb44eec5e6f --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/_global_mmlu_full_nl_stem.yaml @@ -0,0 +1,8 @@ +group: global_mmlu_full_nl_stem +task: + - global_mmlu_full_nl_stem_tasks +aggregate_metric_list: + - metric: acc + weight_by_size: true +metadata: + version: 0.0 diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/_nl_template_yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/_nl_template_yaml new file mode 100644 index 0000000000000000000000000000000000000000..39efbfd177533bc57bdb3c831385a5c386b084e1 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/_nl_template_yaml @@ -0,0 +1,16 @@ +dataset_path: CohereForAI/Global-MMLU +dataset_name: nl +test_split: test +fewshot_split: dev +fewshot_config: + sampler: first_n +output_type: multiple_choice +doc_to_text: "{{question.strip()}}\nA. {{option_a}}\nB. {{option_b}}\nC. {{option_c}}\nD. {{option_d}}\nAnswer:" +doc_to_choice: ["A", "B", "C", "D"] +doc_to_target: answer +metric_list: + - metric: acc + aggregation: mean + higher_is_better: true +metadata: + version: 0.0 diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_abstract_algebra.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_abstract_algebra.yaml new file mode 100644 index 0000000000000000000000000000000000000000..458a36140a17fca9a855c4c1c2852fba05f14b7f --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_abstract_algebra.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _nl_template_yaml +process_docs: !function utils.process_abstract_algebra +tag: global_mmlu_full_nl_stem_tasks +task: global_mmlu_full_nl_abstract_algebra diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_anatomy.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_anatomy.yaml new file mode 100644 index 0000000000000000000000000000000000000000..e4cbd90ef070f7bc12149fb03f46c55d6cbbf591 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_anatomy.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _nl_template_yaml +process_docs: !function utils.process_anatomy +tag: global_mmlu_full_nl_stem_tasks +task: global_mmlu_full_nl_anatomy diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_astronomy.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_astronomy.yaml new file mode 100644 index 0000000000000000000000000000000000000000..84cdf5781ba1c8e50b79069ca291270ab5f742f4 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_astronomy.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _nl_template_yaml +process_docs: !function utils.process_astronomy +tag: global_mmlu_full_nl_stem_tasks +task: global_mmlu_full_nl_astronomy diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_business_ethics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_business_ethics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..f75776f24d231111aa39e224cfeb5e9700b3e253 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_business_ethics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _nl_template_yaml +process_docs: !function utils.process_business_ethics +tag: global_mmlu_full_nl_other_tasks +task: global_mmlu_full_nl_business_ethics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_clinical_knowledge.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_clinical_knowledge.yaml new file mode 100644 index 0000000000000000000000000000000000000000..6e963d0f0d93ae41918a86b2fb5cf63ef243f5cd --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_clinical_knowledge.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _nl_template_yaml +process_docs: !function utils.process_clinical_knowledge +tag: global_mmlu_full_nl_other_tasks +task: global_mmlu_full_nl_clinical_knowledge diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_college_biology.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_college_biology.yaml new file mode 100644 index 0000000000000000000000000000000000000000..e4a3660b07b1681b9f31300e904391ddf57b04fd --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_college_biology.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _nl_template_yaml +process_docs: !function utils.process_college_biology +tag: global_mmlu_full_nl_stem_tasks +task: global_mmlu_full_nl_college_biology diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_college_chemistry.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_college_chemistry.yaml new file mode 100644 index 0000000000000000000000000000000000000000..fa9faed92b3f8c53a714cad2294eb6bd737ca57f --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_college_chemistry.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _nl_template_yaml +process_docs: !function utils.process_college_chemistry +tag: global_mmlu_full_nl_stem_tasks +task: global_mmlu_full_nl_college_chemistry diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_college_computer_science.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_college_computer_science.yaml new file mode 100644 index 0000000000000000000000000000000000000000..b603c309105a1873b5697c7852a15afcafa3d982 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_college_computer_science.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _nl_template_yaml +process_docs: !function utils.process_college_computer_science +tag: global_mmlu_full_nl_stem_tasks +task: global_mmlu_full_nl_college_computer_science diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_college_mathematics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_college_mathematics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..f55207ea49e6588dbfdc36534ed59f514a3d35e1 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_college_mathematics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _nl_template_yaml +process_docs: !function utils.process_college_mathematics +tag: global_mmlu_full_nl_stem_tasks +task: global_mmlu_full_nl_college_mathematics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_college_medicine.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_college_medicine.yaml new file mode 100644 index 0000000000000000000000000000000000000000..5cdda1b7755bcfc2e080b9c5bdd924bdb561c189 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_college_medicine.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _nl_template_yaml +process_docs: !function utils.process_college_medicine +tag: global_mmlu_full_nl_other_tasks +task: global_mmlu_full_nl_college_medicine diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_college_physics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_college_physics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..26d70230fb8b945e0554ff7df68e0b4e5fb4e114 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_college_physics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _nl_template_yaml +process_docs: !function utils.process_college_physics +tag: global_mmlu_full_nl_stem_tasks +task: global_mmlu_full_nl_college_physics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_computer_security.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_computer_security.yaml new file mode 100644 index 0000000000000000000000000000000000000000..01a8a74742e44468220cf1a7f85aa68f1d53fa13 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_computer_security.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _nl_template_yaml +process_docs: !function utils.process_computer_security +tag: global_mmlu_full_nl_stem_tasks +task: global_mmlu_full_nl_computer_security diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_conceptual_physics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_conceptual_physics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..cccd266651a36b14d85ece24d3d639964bd7106d --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_conceptual_physics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _nl_template_yaml +process_docs: !function utils.process_conceptual_physics +tag: global_mmlu_full_nl_stem_tasks +task: global_mmlu_full_nl_conceptual_physics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_econometrics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_econometrics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..22ad59bd462ceee61f12955852634723b02852a1 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_econometrics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _nl_template_yaml +process_docs: !function utils.process_econometrics +tag: global_mmlu_full_nl_social_sciences_tasks +task: global_mmlu_full_nl_econometrics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_electrical_engineering.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_electrical_engineering.yaml new file mode 100644 index 0000000000000000000000000000000000000000..3aca226fdff9fe7219463f5dcc2cc41bb4fa16c1 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_electrical_engineering.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _nl_template_yaml +process_docs: !function utils.process_electrical_engineering +tag: global_mmlu_full_nl_stem_tasks +task: global_mmlu_full_nl_electrical_engineering diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_elementary_mathematics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_elementary_mathematics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..2118a1d5d6935b9ba7603b393babba2a027536c4 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_elementary_mathematics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _nl_template_yaml +process_docs: !function utils.process_elementary_mathematics +tag: global_mmlu_full_nl_stem_tasks +task: global_mmlu_full_nl_elementary_mathematics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_formal_logic.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_formal_logic.yaml new file mode 100644 index 0000000000000000000000000000000000000000..5fd86105df589b16c08f9c8abfdfc7eca6cbf712 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_formal_logic.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _nl_template_yaml +process_docs: !function utils.process_formal_logic +tag: global_mmlu_full_nl_humanities_tasks +task: global_mmlu_full_nl_formal_logic diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_global_facts.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_global_facts.yaml new file mode 100644 index 0000000000000000000000000000000000000000..d7147d517cfcfd02013b3afbf8ba0ec11e95d180 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_global_facts.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _nl_template_yaml +process_docs: !function utils.process_global_facts +tag: global_mmlu_full_nl_other_tasks +task: global_mmlu_full_nl_global_facts diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_high_school_biology.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_high_school_biology.yaml new file mode 100644 index 0000000000000000000000000000000000000000..271b54f65e95485e1fdfb3b4e3d39d282f6c5f1b --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_high_school_biology.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _nl_template_yaml +process_docs: !function utils.process_high_school_biology +tag: global_mmlu_full_nl_stem_tasks +task: global_mmlu_full_nl_high_school_biology diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_high_school_chemistry.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_high_school_chemistry.yaml new file mode 100644 index 0000000000000000000000000000000000000000..921abd1745e94ee09467c2d63f0575455fe1d750 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_high_school_chemistry.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _nl_template_yaml +process_docs: !function utils.process_high_school_chemistry +tag: global_mmlu_full_nl_stem_tasks +task: global_mmlu_full_nl_high_school_chemistry diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_high_school_computer_science.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_high_school_computer_science.yaml new file mode 100644 index 0000000000000000000000000000000000000000..ea190bea8face93cccc33b337750bb39707ac8f5 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_high_school_computer_science.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _nl_template_yaml +process_docs: !function utils.process_high_school_computer_science +tag: global_mmlu_full_nl_stem_tasks +task: global_mmlu_full_nl_high_school_computer_science diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_high_school_european_history.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_high_school_european_history.yaml new file mode 100644 index 0000000000000000000000000000000000000000..c348d482a074bcc6dba6eddef8905a8df859cd8c --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_high_school_european_history.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _nl_template_yaml +process_docs: !function utils.process_high_school_european_history +tag: global_mmlu_full_nl_humanities_tasks +task: global_mmlu_full_nl_high_school_european_history diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_high_school_geography.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_high_school_geography.yaml new file mode 100644 index 0000000000000000000000000000000000000000..de31a63b8df61eee7f08b5a96e283c7e5b93d400 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_high_school_geography.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _nl_template_yaml +process_docs: !function utils.process_high_school_geography +tag: global_mmlu_full_nl_social_sciences_tasks +task: global_mmlu_full_nl_high_school_geography diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_high_school_government_and_politics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_high_school_government_and_politics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..bc0e3cb1bc79dc6b5400497c07ba789b72cec23e --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_high_school_government_and_politics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _nl_template_yaml +process_docs: !function utils.process_high_school_government_and_politics +tag: global_mmlu_full_nl_social_sciences_tasks +task: global_mmlu_full_nl_high_school_government_and_politics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_high_school_macroeconomics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_high_school_macroeconomics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..2e221c685fb2d15b7d9719e2e17d25ec44b70c1e --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_high_school_macroeconomics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _nl_template_yaml +process_docs: !function utils.process_high_school_macroeconomics +tag: global_mmlu_full_nl_social_sciences_tasks +task: global_mmlu_full_nl_high_school_macroeconomics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_high_school_mathematics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_high_school_mathematics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..137158a690d175135c0ebf2d94f038ef54fbf615 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_high_school_mathematics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _nl_template_yaml +process_docs: !function utils.process_high_school_mathematics +tag: global_mmlu_full_nl_stem_tasks +task: global_mmlu_full_nl_high_school_mathematics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_high_school_microeconomics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_high_school_microeconomics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..27b426c06f46f522fe70dccf7241f05b27af505a --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_high_school_microeconomics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _nl_template_yaml +process_docs: !function utils.process_high_school_microeconomics +tag: global_mmlu_full_nl_social_sciences_tasks +task: global_mmlu_full_nl_high_school_microeconomics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_high_school_physics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_high_school_physics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..746df49e81b176db860a12d31781a06ae6e93448 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_high_school_physics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _nl_template_yaml +process_docs: !function utils.process_high_school_physics +tag: global_mmlu_full_nl_stem_tasks +task: global_mmlu_full_nl_high_school_physics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_high_school_psychology.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_high_school_psychology.yaml new file mode 100644 index 0000000000000000000000000000000000000000..89cb42d9472a8e6517a6eb5de45312ecb826e02a --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_high_school_psychology.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _nl_template_yaml +process_docs: !function utils.process_high_school_psychology +tag: global_mmlu_full_nl_social_sciences_tasks +task: global_mmlu_full_nl_high_school_psychology diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_high_school_statistics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_high_school_statistics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..e27082c4cb8c320b8dab7e61d92db36c46cf1484 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_high_school_statistics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _nl_template_yaml +process_docs: !function utils.process_high_school_statistics +tag: global_mmlu_full_nl_stem_tasks +task: global_mmlu_full_nl_high_school_statistics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_high_school_us_history.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_high_school_us_history.yaml new file mode 100644 index 0000000000000000000000000000000000000000..66efc58c9673f3afb5129caa39cd06fe31e275e0 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_high_school_us_history.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _nl_template_yaml +process_docs: !function utils.process_high_school_us_history +tag: global_mmlu_full_nl_humanities_tasks +task: global_mmlu_full_nl_high_school_us_history diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_high_school_world_history.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_high_school_world_history.yaml new file mode 100644 index 0000000000000000000000000000000000000000..83b653454773bf44e5a8fe8ad35125fa68b120de --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_high_school_world_history.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _nl_template_yaml +process_docs: !function utils.process_high_school_world_history +tag: global_mmlu_full_nl_humanities_tasks +task: global_mmlu_full_nl_high_school_world_history diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_human_aging.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_human_aging.yaml new file mode 100644 index 0000000000000000000000000000000000000000..82e00b4b8fa93e31a3307d40482e1dc628f6a16e --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_human_aging.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _nl_template_yaml +process_docs: !function utils.process_human_aging +tag: global_mmlu_full_nl_other_tasks +task: global_mmlu_full_nl_human_aging diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_human_sexuality.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_human_sexuality.yaml new file mode 100644 index 0000000000000000000000000000000000000000..468589da275b9236dbb72c6eaee686c5c9fdcdc2 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_human_sexuality.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _nl_template_yaml +process_docs: !function utils.process_human_sexuality +tag: global_mmlu_full_nl_social_sciences_tasks +task: global_mmlu_full_nl_human_sexuality diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_international_law.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_international_law.yaml new file mode 100644 index 0000000000000000000000000000000000000000..e5bf62a98e7b3442f821bc3209d124e061e49052 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_international_law.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _nl_template_yaml +process_docs: !function utils.process_international_law +tag: global_mmlu_full_nl_humanities_tasks +task: global_mmlu_full_nl_international_law diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_jurisprudence.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_jurisprudence.yaml new file mode 100644 index 0000000000000000000000000000000000000000..7b53361381a6c63cf3eb4d3a0bdace7c7a29a10c --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_jurisprudence.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _nl_template_yaml +process_docs: !function utils.process_jurisprudence +tag: global_mmlu_full_nl_humanities_tasks +task: global_mmlu_full_nl_jurisprudence diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_logical_fallacies.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_logical_fallacies.yaml new file mode 100644 index 0000000000000000000000000000000000000000..de862b6631b9749b0f6f343c99f35e7f8d2b6fc4 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_logical_fallacies.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _nl_template_yaml +process_docs: !function utils.process_logical_fallacies +tag: global_mmlu_full_nl_humanities_tasks +task: global_mmlu_full_nl_logical_fallacies diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_machine_learning.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_machine_learning.yaml new file mode 100644 index 0000000000000000000000000000000000000000..c205af001d210fdc47806f322b6771ca24f900c9 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_machine_learning.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _nl_template_yaml +process_docs: !function utils.process_machine_learning +tag: global_mmlu_full_nl_stem_tasks +task: global_mmlu_full_nl_machine_learning diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_management.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_management.yaml new file mode 100644 index 0000000000000000000000000000000000000000..5b624af8bf29f0966c7d9d5e816c3fffb704f743 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_management.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _nl_template_yaml +process_docs: !function utils.process_management +tag: global_mmlu_full_nl_other_tasks +task: global_mmlu_full_nl_management diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_marketing.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_marketing.yaml new file mode 100644 index 0000000000000000000000000000000000000000..81658e9f0278dc681e525ca52f3abd6bc8f5d657 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_marketing.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _nl_template_yaml +process_docs: !function utils.process_marketing +tag: global_mmlu_full_nl_other_tasks +task: global_mmlu_full_nl_marketing diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_medical_genetics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_medical_genetics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..f8e52c0d4ec5d2231ea89155dde624887acbd6f1 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_medical_genetics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _nl_template_yaml +process_docs: !function utils.process_medical_genetics +tag: global_mmlu_full_nl_other_tasks +task: global_mmlu_full_nl_medical_genetics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_miscellaneous.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_miscellaneous.yaml new file mode 100644 index 0000000000000000000000000000000000000000..31af482ec8ca1a1bbd97cd82f07651b2a35b092d --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_miscellaneous.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _nl_template_yaml +process_docs: !function utils.process_miscellaneous +tag: global_mmlu_full_nl_other_tasks +task: global_mmlu_full_nl_miscellaneous diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_moral_disputes.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_moral_disputes.yaml new file mode 100644 index 0000000000000000000000000000000000000000..853de0c12fb9416a94ce8e8e1bd04e3a63d2ffa8 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_moral_disputes.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _nl_template_yaml +process_docs: !function utils.process_moral_disputes +tag: global_mmlu_full_nl_humanities_tasks +task: global_mmlu_full_nl_moral_disputes diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_moral_scenarios.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_moral_scenarios.yaml new file mode 100644 index 0000000000000000000000000000000000000000..8b86e04500b404fa145fe1bbaa70ac9e36bdc6d7 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_moral_scenarios.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _nl_template_yaml +process_docs: !function utils.process_moral_scenarios +tag: global_mmlu_full_nl_humanities_tasks +task: global_mmlu_full_nl_moral_scenarios diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_nutrition.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_nutrition.yaml new file mode 100644 index 0000000000000000000000000000000000000000..96036dae3809fbdfe6187de08a48d666e96245c1 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_nutrition.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _nl_template_yaml +process_docs: !function utils.process_nutrition +tag: global_mmlu_full_nl_other_tasks +task: global_mmlu_full_nl_nutrition diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_philosophy.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_philosophy.yaml new file mode 100644 index 0000000000000000000000000000000000000000..84e827dd19b3e6184f0868c7f26bd1e38eddc420 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_philosophy.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _nl_template_yaml +process_docs: !function utils.process_philosophy +tag: global_mmlu_full_nl_humanities_tasks +task: global_mmlu_full_nl_philosophy diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_prehistory.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_prehistory.yaml new file mode 100644 index 0000000000000000000000000000000000000000..f49c8a5e84786200acbb4f917b8a697f936aa1a9 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_prehistory.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _nl_template_yaml +process_docs: !function utils.process_prehistory +tag: global_mmlu_full_nl_humanities_tasks +task: global_mmlu_full_nl_prehistory diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_professional_accounting.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_professional_accounting.yaml new file mode 100644 index 0000000000000000000000000000000000000000..45484116ddcb28ead1fd09b72aa7c169faa5c0b3 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_professional_accounting.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _nl_template_yaml +process_docs: !function utils.process_professional_accounting +tag: global_mmlu_full_nl_other_tasks +task: global_mmlu_full_nl_professional_accounting diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_professional_law.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_professional_law.yaml new file mode 100644 index 0000000000000000000000000000000000000000..17b28cd32c70f580fe1d6e80aa8dc82d6a4c6875 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_professional_law.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _nl_template_yaml +process_docs: !function utils.process_professional_law +tag: global_mmlu_full_nl_humanities_tasks +task: global_mmlu_full_nl_professional_law diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_professional_medicine.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_professional_medicine.yaml new file mode 100644 index 0000000000000000000000000000000000000000..f4db01bc67184c4ece2b11d64a38f8d3ddc8b441 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_professional_medicine.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _nl_template_yaml +process_docs: !function utils.process_professional_medicine +tag: global_mmlu_full_nl_other_tasks +task: global_mmlu_full_nl_professional_medicine diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_professional_psychology.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_professional_psychology.yaml new file mode 100644 index 0000000000000000000000000000000000000000..be586b45d96457457c2a3ecc27cc8dbc06ce4f26 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_professional_psychology.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _nl_template_yaml +process_docs: !function utils.process_professional_psychology +tag: global_mmlu_full_nl_social_sciences_tasks +task: global_mmlu_full_nl_professional_psychology diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_public_relations.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_public_relations.yaml new file mode 100644 index 0000000000000000000000000000000000000000..2ffe584887a64f84ce5ec1a34c9a83e9c7b8f1bd --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_public_relations.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _nl_template_yaml +process_docs: !function utils.process_public_relations +tag: global_mmlu_full_nl_social_sciences_tasks +task: global_mmlu_full_nl_public_relations diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_security_studies.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_security_studies.yaml new file mode 100644 index 0000000000000000000000000000000000000000..b6c76948b5866e83a1070fb091a3d2fd2df8b5ec --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_security_studies.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _nl_template_yaml +process_docs: !function utils.process_security_studies +tag: global_mmlu_full_nl_social_sciences_tasks +task: global_mmlu_full_nl_security_studies diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_sociology.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_sociology.yaml new file mode 100644 index 0000000000000000000000000000000000000000..983e13cde755f5c09c4c6338bf463a1ffd89892a --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_sociology.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _nl_template_yaml +process_docs: !function utils.process_sociology +tag: global_mmlu_full_nl_social_sciences_tasks +task: global_mmlu_full_nl_sociology diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_us_foreign_policy.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_us_foreign_policy.yaml new file mode 100644 index 0000000000000000000000000000000000000000..bd6b6227bf6c656c75e4f7d35991761c92e72a19 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_us_foreign_policy.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _nl_template_yaml +process_docs: !function utils.process_us_foreign_policy +tag: global_mmlu_full_nl_social_sciences_tasks +task: global_mmlu_full_nl_us_foreign_policy diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_virology.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_virology.yaml new file mode 100644 index 0000000000000000000000000000000000000000..92d1973bb57dc8c7f2a57f8d0d14547b55229041 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_virology.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _nl_template_yaml +process_docs: !function utils.process_virology +tag: global_mmlu_full_nl_other_tasks +task: global_mmlu_full_nl_virology diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_world_religions.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_world_religions.yaml new file mode 100644 index 0000000000000000000000000000000000000000..a8c2eccaa24955da352770a507f00e59fca5dfe6 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/global_mmlu_full_nl_world_religions.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _nl_template_yaml +process_docs: !function utils.process_world_religions +tag: global_mmlu_full_nl_humanities_tasks +task: global_mmlu_full_nl_world_religions diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/utils.py b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/utils.py new file mode 100644 index 0000000000000000000000000000000000000000..7df72cb061f0fecba46e15ff9b57552817979afb --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/nl/utils.py @@ -0,0 +1,73 @@ +from functools import partial + + +SUBJECTS = [ + "abstract_algebra", + "anatomy", + "astronomy", + "business_ethics", + "clinical_knowledge", + "college_biology", + "college_chemistry", + "college_computer_science", + "college_mathematics", + "college_medicine", + "college_physics", + "computer_security", + "conceptual_physics", + "econometrics", + "electrical_engineering", + "elementary_mathematics", + "formal_logic", + "global_facts", + "high_school_biology", + "high_school_chemistry", + "high_school_computer_science", + "high_school_european_history", + "high_school_geography", + "high_school_government_and_politics", + "high_school_macroeconomics", + "high_school_mathematics", + "high_school_microeconomics", + "high_school_physics", + "high_school_psychology", + "high_school_statistics", + "high_school_us_history", + "high_school_world_history", + "human_aging", + "human_sexuality", + "international_law", + "jurisprudence", + "logical_fallacies", + "machine_learning", + "management", + "marketing", + "medical_genetics", + "miscellaneous", + "moral_disputes", + "moral_scenarios", + "nutrition", + "philosophy", + "prehistory", + "professional_accounting", + "professional_law", + "professional_medicine", + "professional_psychology", + "public_relations", + "security_studies", + "sociology", + "us_foreign_policy", + "virology", + "world_religions", +] + + +def process_docs(dataset, subject): + return dataset.filter(lambda x: x["subject"] == subject) + + +process_functions = { + f"process_{subject}": partial(process_docs, subject=subject) for subject in SUBJECTS +} + +globals().update(process_functions) diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/_global_mmlu_full_ny.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/_global_mmlu_full_ny.yaml new file mode 100644 index 0000000000000000000000000000000000000000..80739c2c70a5aceb71209a0851f53cc7ce3fc500 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/_global_mmlu_full_ny.yaml @@ -0,0 +1,11 @@ +group: global_mmlu_full_ny +task: + - global_mmlu_full_ny_stem + - global_mmlu_full_ny_other + - global_mmlu_full_ny_social_sciences + - global_mmlu_full_ny_humanities +aggregate_metric_list: + - metric: acc + weight_by_size: True +metadata: + version: 0.0 diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/_global_mmlu_full_ny_humanities.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/_global_mmlu_full_ny_humanities.yaml new file mode 100644 index 0000000000000000000000000000000000000000..89e7618fbea41e79405b24889b788bf59d268fc4 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/_global_mmlu_full_ny_humanities.yaml @@ -0,0 +1,8 @@ +group: global_mmlu_full_ny_humanities +task: + - global_mmlu_full_ny_humanities_tasks +aggregate_metric_list: + - metric: acc + weight_by_size: true +metadata: + version: 0.0 diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/_global_mmlu_full_ny_other.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/_global_mmlu_full_ny_other.yaml new file mode 100644 index 0000000000000000000000000000000000000000..51b90446d1fd2362581c3ab9fb159e3fe44d8ab7 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/_global_mmlu_full_ny_other.yaml @@ -0,0 +1,8 @@ +group: global_mmlu_full_ny_other +task: + - global_mmlu_full_ny_other_tasks +aggregate_metric_list: + - metric: acc + weight_by_size: true +metadata: + version: 0.0 diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/_global_mmlu_full_ny_social_sciences.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/_global_mmlu_full_ny_social_sciences.yaml new file mode 100644 index 0000000000000000000000000000000000000000..b711dfdf14029a040e965c36fdae79e0702c31e7 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/_global_mmlu_full_ny_social_sciences.yaml @@ -0,0 +1,8 @@ +group: global_mmlu_full_ny_social_sciences +task: + - global_mmlu_full_ny_social_sciences_tasks +aggregate_metric_list: + - metric: acc + weight_by_size: true +metadata: + version: 0.0 diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/_global_mmlu_full_ny_stem.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/_global_mmlu_full_ny_stem.yaml new file mode 100644 index 0000000000000000000000000000000000000000..99bf9d95df56a763b2bca25c05d20b1b4f2ba366 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/_global_mmlu_full_ny_stem.yaml @@ -0,0 +1,8 @@ +group: global_mmlu_full_ny_stem +task: + - global_mmlu_full_ny_stem_tasks +aggregate_metric_list: + - metric: acc + weight_by_size: true +metadata: + version: 0.0 diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/_ny_template_yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/_ny_template_yaml new file mode 100644 index 0000000000000000000000000000000000000000..069a9446ec733541d3295268f7f6a943f13435a8 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/_ny_template_yaml @@ -0,0 +1,16 @@ +dataset_path: CohereForAI/Global-MMLU +dataset_name: ny +test_split: test +fewshot_split: dev +fewshot_config: + sampler: first_n +output_type: multiple_choice +doc_to_text: "{{question.strip()}}\nA. {{option_a}}\nB. {{option_b}}\nC. {{option_c}}\nD. {{option_d}}\nAnswer:" +doc_to_choice: ["A", "B", "C", "D"] +doc_to_target: answer +metric_list: + - metric: acc + aggregation: mean + higher_is_better: true +metadata: + version: 0.0 diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_abstract_algebra.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_abstract_algebra.yaml new file mode 100644 index 0000000000000000000000000000000000000000..2e3d7c33b7cd893564177664390bc62554a38e08 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_abstract_algebra.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ny_template_yaml +process_docs: !function utils.process_abstract_algebra +tag: global_mmlu_full_ny_stem_tasks +task: global_mmlu_full_ny_abstract_algebra diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_anatomy.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_anatomy.yaml new file mode 100644 index 0000000000000000000000000000000000000000..60806afc61e6e35998b3c96d2631b01b66121923 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_anatomy.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ny_template_yaml +process_docs: !function utils.process_anatomy +tag: global_mmlu_full_ny_stem_tasks +task: global_mmlu_full_ny_anatomy diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_astronomy.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_astronomy.yaml new file mode 100644 index 0000000000000000000000000000000000000000..afbcb482fb44bac4274140232adcb014dac97fd8 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_astronomy.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ny_template_yaml +process_docs: !function utils.process_astronomy +tag: global_mmlu_full_ny_stem_tasks +task: global_mmlu_full_ny_astronomy diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_business_ethics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_business_ethics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..6f8981bd0280c4743b0461ee2b0d8587c8bc0b01 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_business_ethics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ny_template_yaml +process_docs: !function utils.process_business_ethics +tag: global_mmlu_full_ny_other_tasks +task: global_mmlu_full_ny_business_ethics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_clinical_knowledge.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_clinical_knowledge.yaml new file mode 100644 index 0000000000000000000000000000000000000000..ff44dd679f7e478f3663c7c8ad0aa1b37afdedef --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_clinical_knowledge.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ny_template_yaml +process_docs: !function utils.process_clinical_knowledge +tag: global_mmlu_full_ny_other_tasks +task: global_mmlu_full_ny_clinical_knowledge diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_college_biology.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_college_biology.yaml new file mode 100644 index 0000000000000000000000000000000000000000..da5ce37080d7665fe15fcfc067f40dfdab2a37b4 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_college_biology.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ny_template_yaml +process_docs: !function utils.process_college_biology +tag: global_mmlu_full_ny_stem_tasks +task: global_mmlu_full_ny_college_biology diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_college_chemistry.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_college_chemistry.yaml new file mode 100644 index 0000000000000000000000000000000000000000..d62bce830bd1a21fe0d05839cffc53d17f938061 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_college_chemistry.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ny_template_yaml +process_docs: !function utils.process_college_chemistry +tag: global_mmlu_full_ny_stem_tasks +task: global_mmlu_full_ny_college_chemistry diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_college_computer_science.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_college_computer_science.yaml new file mode 100644 index 0000000000000000000000000000000000000000..48cd98d55bd649d2b3a94587b17ab59ad9d6f71f --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_college_computer_science.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ny_template_yaml +process_docs: !function utils.process_college_computer_science +tag: global_mmlu_full_ny_stem_tasks +task: global_mmlu_full_ny_college_computer_science diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_college_mathematics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_college_mathematics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..ed77ba9c18df83ce3a87f9e6d1aca93974d6f35e --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_college_mathematics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ny_template_yaml +process_docs: !function utils.process_college_mathematics +tag: global_mmlu_full_ny_stem_tasks +task: global_mmlu_full_ny_college_mathematics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_college_medicine.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_college_medicine.yaml new file mode 100644 index 0000000000000000000000000000000000000000..9cd8aa2f43fd9f29c99d1afc2b722f613f801596 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_college_medicine.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ny_template_yaml +process_docs: !function utils.process_college_medicine +tag: global_mmlu_full_ny_other_tasks +task: global_mmlu_full_ny_college_medicine diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_college_physics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_college_physics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..66d5dc27b76d8d0e670244fa3d1ddf5003726b9c --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_college_physics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ny_template_yaml +process_docs: !function utils.process_college_physics +tag: global_mmlu_full_ny_stem_tasks +task: global_mmlu_full_ny_college_physics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_computer_security.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_computer_security.yaml new file mode 100644 index 0000000000000000000000000000000000000000..8a9dae6225be59b56d6ae32e516846f574bcdd9f --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_computer_security.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ny_template_yaml +process_docs: !function utils.process_computer_security +tag: global_mmlu_full_ny_stem_tasks +task: global_mmlu_full_ny_computer_security diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_conceptual_physics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_conceptual_physics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..8d160ffc658171e9aff349f431ba3c3270ecdbf9 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_conceptual_physics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ny_template_yaml +process_docs: !function utils.process_conceptual_physics +tag: global_mmlu_full_ny_stem_tasks +task: global_mmlu_full_ny_conceptual_physics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_econometrics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_econometrics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..88af709a618435ea07ae6acc136471a593bc7e91 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_econometrics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ny_template_yaml +process_docs: !function utils.process_econometrics +tag: global_mmlu_full_ny_social_sciences_tasks +task: global_mmlu_full_ny_econometrics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_electrical_engineering.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_electrical_engineering.yaml new file mode 100644 index 0000000000000000000000000000000000000000..d835f1e003fb6aabba6b787c322667351924a6f4 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_electrical_engineering.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ny_template_yaml +process_docs: !function utils.process_electrical_engineering +tag: global_mmlu_full_ny_stem_tasks +task: global_mmlu_full_ny_electrical_engineering diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_elementary_mathematics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_elementary_mathematics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..558ffd0bac03a9fab192fd1e7c581841acf8c4e2 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_elementary_mathematics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ny_template_yaml +process_docs: !function utils.process_elementary_mathematics +tag: global_mmlu_full_ny_stem_tasks +task: global_mmlu_full_ny_elementary_mathematics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_formal_logic.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_formal_logic.yaml new file mode 100644 index 0000000000000000000000000000000000000000..cce0df19215c36e911fbe4b8757851da6dc49029 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_formal_logic.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ny_template_yaml +process_docs: !function utils.process_formal_logic +tag: global_mmlu_full_ny_humanities_tasks +task: global_mmlu_full_ny_formal_logic diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_global_facts.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_global_facts.yaml new file mode 100644 index 0000000000000000000000000000000000000000..6ce027a55f9e74dda90a9028259213afa97717c5 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_global_facts.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ny_template_yaml +process_docs: !function utils.process_global_facts +tag: global_mmlu_full_ny_other_tasks +task: global_mmlu_full_ny_global_facts diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_high_school_biology.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_high_school_biology.yaml new file mode 100644 index 0000000000000000000000000000000000000000..a729008d55fc6da1ac8f681616cb881a13062a66 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_high_school_biology.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ny_template_yaml +process_docs: !function utils.process_high_school_biology +tag: global_mmlu_full_ny_stem_tasks +task: global_mmlu_full_ny_high_school_biology diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_high_school_chemistry.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_high_school_chemistry.yaml new file mode 100644 index 0000000000000000000000000000000000000000..79771bfb651ac6bce183fea43fac60366d28530d --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_high_school_chemistry.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ny_template_yaml +process_docs: !function utils.process_high_school_chemistry +tag: global_mmlu_full_ny_stem_tasks +task: global_mmlu_full_ny_high_school_chemistry diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_high_school_computer_science.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_high_school_computer_science.yaml new file mode 100644 index 0000000000000000000000000000000000000000..6889806f2bf0e1cac4de62afec43278bbb9caad3 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_high_school_computer_science.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ny_template_yaml +process_docs: !function utils.process_high_school_computer_science +tag: global_mmlu_full_ny_stem_tasks +task: global_mmlu_full_ny_high_school_computer_science diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_high_school_european_history.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_high_school_european_history.yaml new file mode 100644 index 0000000000000000000000000000000000000000..29e6e4a5a3cf39c64c0bda8413c91111d1238416 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_high_school_european_history.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ny_template_yaml +process_docs: !function utils.process_high_school_european_history +tag: global_mmlu_full_ny_humanities_tasks +task: global_mmlu_full_ny_high_school_european_history diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_high_school_geography.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_high_school_geography.yaml new file mode 100644 index 0000000000000000000000000000000000000000..447db75fa9d2f3cf4401e576f824eab474235606 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_high_school_geography.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ny_template_yaml +process_docs: !function utils.process_high_school_geography +tag: global_mmlu_full_ny_social_sciences_tasks +task: global_mmlu_full_ny_high_school_geography diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_high_school_government_and_politics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_high_school_government_and_politics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..e543cf7607c15f25b03b4bbf1f2e78b1a5842d02 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_high_school_government_and_politics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ny_template_yaml +process_docs: !function utils.process_high_school_government_and_politics +tag: global_mmlu_full_ny_social_sciences_tasks +task: global_mmlu_full_ny_high_school_government_and_politics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_high_school_macroeconomics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_high_school_macroeconomics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..61c49e750f738e195c4152b3d3d89995ed8139ae --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_high_school_macroeconomics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ny_template_yaml +process_docs: !function utils.process_high_school_macroeconomics +tag: global_mmlu_full_ny_social_sciences_tasks +task: global_mmlu_full_ny_high_school_macroeconomics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_high_school_mathematics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_high_school_mathematics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..db228d02198cf9967dc51d9bb570eb70422f4924 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_high_school_mathematics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ny_template_yaml +process_docs: !function utils.process_high_school_mathematics +tag: global_mmlu_full_ny_stem_tasks +task: global_mmlu_full_ny_high_school_mathematics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_high_school_microeconomics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_high_school_microeconomics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..62d87c862e8a57c682b3af042f56af6219371a05 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_high_school_microeconomics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ny_template_yaml +process_docs: !function utils.process_high_school_microeconomics +tag: global_mmlu_full_ny_social_sciences_tasks +task: global_mmlu_full_ny_high_school_microeconomics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_high_school_physics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_high_school_physics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..54c15d66e5659537fdd0aba7eed741dcd5fd538e --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_high_school_physics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ny_template_yaml +process_docs: !function utils.process_high_school_physics +tag: global_mmlu_full_ny_stem_tasks +task: global_mmlu_full_ny_high_school_physics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_high_school_psychology.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_high_school_psychology.yaml new file mode 100644 index 0000000000000000000000000000000000000000..4f7d8b5a1c730bdb1131db0e92a078061e0112c4 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_high_school_psychology.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ny_template_yaml +process_docs: !function utils.process_high_school_psychology +tag: global_mmlu_full_ny_social_sciences_tasks +task: global_mmlu_full_ny_high_school_psychology diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_high_school_statistics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_high_school_statistics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..f53235b878aec25fb7ed3608f8cb2c2ee55852f5 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_high_school_statistics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ny_template_yaml +process_docs: !function utils.process_high_school_statistics +tag: global_mmlu_full_ny_stem_tasks +task: global_mmlu_full_ny_high_school_statistics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_high_school_us_history.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_high_school_us_history.yaml new file mode 100644 index 0000000000000000000000000000000000000000..1d413b985b68b658efc8d4771f9771a153895dd0 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_high_school_us_history.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ny_template_yaml +process_docs: !function utils.process_high_school_us_history +tag: global_mmlu_full_ny_humanities_tasks +task: global_mmlu_full_ny_high_school_us_history diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_high_school_world_history.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_high_school_world_history.yaml new file mode 100644 index 0000000000000000000000000000000000000000..4adf2e8b12c8d663e7941079d718ca04d7bd0ca1 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_high_school_world_history.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ny_template_yaml +process_docs: !function utils.process_high_school_world_history +tag: global_mmlu_full_ny_humanities_tasks +task: global_mmlu_full_ny_high_school_world_history diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_human_aging.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_human_aging.yaml new file mode 100644 index 0000000000000000000000000000000000000000..9660b7b1308547c83102e1e7a9f6a603c4889f7d --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_human_aging.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ny_template_yaml +process_docs: !function utils.process_human_aging +tag: global_mmlu_full_ny_other_tasks +task: global_mmlu_full_ny_human_aging diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_human_sexuality.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_human_sexuality.yaml new file mode 100644 index 0000000000000000000000000000000000000000..11a6f2d4f660c23029df81b57a7914f193b86098 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_human_sexuality.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ny_template_yaml +process_docs: !function utils.process_human_sexuality +tag: global_mmlu_full_ny_social_sciences_tasks +task: global_mmlu_full_ny_human_sexuality diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_international_law.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_international_law.yaml new file mode 100644 index 0000000000000000000000000000000000000000..9a46ff6a9c906e7538795d0494bb7368f2708315 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_international_law.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ny_template_yaml +process_docs: !function utils.process_international_law +tag: global_mmlu_full_ny_humanities_tasks +task: global_mmlu_full_ny_international_law diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_jurisprudence.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_jurisprudence.yaml new file mode 100644 index 0000000000000000000000000000000000000000..e4606df53bddfad169bc678e41cc6e97447198cc --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_jurisprudence.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ny_template_yaml +process_docs: !function utils.process_jurisprudence +tag: global_mmlu_full_ny_humanities_tasks +task: global_mmlu_full_ny_jurisprudence diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_logical_fallacies.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_logical_fallacies.yaml new file mode 100644 index 0000000000000000000000000000000000000000..6edade038f8ddbe8bb8dfc37b9be4ffdd620cca4 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_logical_fallacies.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ny_template_yaml +process_docs: !function utils.process_logical_fallacies +tag: global_mmlu_full_ny_humanities_tasks +task: global_mmlu_full_ny_logical_fallacies diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_machine_learning.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_machine_learning.yaml new file mode 100644 index 0000000000000000000000000000000000000000..765b2201444e56f44f6447627e8edd57dde558d6 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_machine_learning.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ny_template_yaml +process_docs: !function utils.process_machine_learning +tag: global_mmlu_full_ny_stem_tasks +task: global_mmlu_full_ny_machine_learning diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_management.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_management.yaml new file mode 100644 index 0000000000000000000000000000000000000000..a699a70dbada19b4a2c32c8fc4d554c34a23b4eb --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_management.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ny_template_yaml +process_docs: !function utils.process_management +tag: global_mmlu_full_ny_other_tasks +task: global_mmlu_full_ny_management diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_marketing.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_marketing.yaml new file mode 100644 index 0000000000000000000000000000000000000000..596d69375d64438f70093658783ae167ca236d0a --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_marketing.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ny_template_yaml +process_docs: !function utils.process_marketing +tag: global_mmlu_full_ny_other_tasks +task: global_mmlu_full_ny_marketing diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_medical_genetics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_medical_genetics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..4fae66a7c5ecef3de82fef9fb280c3a3c2c3ca14 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_medical_genetics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ny_template_yaml +process_docs: !function utils.process_medical_genetics +tag: global_mmlu_full_ny_other_tasks +task: global_mmlu_full_ny_medical_genetics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_miscellaneous.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_miscellaneous.yaml new file mode 100644 index 0000000000000000000000000000000000000000..8555e173ab35f1179371bfd767cb28888651cb34 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_miscellaneous.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ny_template_yaml +process_docs: !function utils.process_miscellaneous +tag: global_mmlu_full_ny_other_tasks +task: global_mmlu_full_ny_miscellaneous diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_moral_disputes.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_moral_disputes.yaml new file mode 100644 index 0000000000000000000000000000000000000000..b64f4d9da4eddbfaa56c9318b1aa68ba17deb859 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_moral_disputes.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ny_template_yaml +process_docs: !function utils.process_moral_disputes +tag: global_mmlu_full_ny_humanities_tasks +task: global_mmlu_full_ny_moral_disputes diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_moral_scenarios.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_moral_scenarios.yaml new file mode 100644 index 0000000000000000000000000000000000000000..c73f9f1af706ecde712b11fc5ddcd4792a2c919a --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_moral_scenarios.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ny_template_yaml +process_docs: !function utils.process_moral_scenarios +tag: global_mmlu_full_ny_humanities_tasks +task: global_mmlu_full_ny_moral_scenarios diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_nutrition.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_nutrition.yaml new file mode 100644 index 0000000000000000000000000000000000000000..456f4cb63edfc56346a4077b320ad77ff9c882cd --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_nutrition.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ny_template_yaml +process_docs: !function utils.process_nutrition +tag: global_mmlu_full_ny_other_tasks +task: global_mmlu_full_ny_nutrition diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_philosophy.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_philosophy.yaml new file mode 100644 index 0000000000000000000000000000000000000000..d0e0e05edddd27e5260b4890837d380c2b1206a5 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_philosophy.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ny_template_yaml +process_docs: !function utils.process_philosophy +tag: global_mmlu_full_ny_humanities_tasks +task: global_mmlu_full_ny_philosophy diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_prehistory.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_prehistory.yaml new file mode 100644 index 0000000000000000000000000000000000000000..d65c6be1e2469f956e658ee2e1ea4caa58ee09df --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_prehistory.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ny_template_yaml +process_docs: !function utils.process_prehistory +tag: global_mmlu_full_ny_humanities_tasks +task: global_mmlu_full_ny_prehistory diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_professional_accounting.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_professional_accounting.yaml new file mode 100644 index 0000000000000000000000000000000000000000..c152c80efd12264f17a001672b23329a0ae85379 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_professional_accounting.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ny_template_yaml +process_docs: !function utils.process_professional_accounting +tag: global_mmlu_full_ny_other_tasks +task: global_mmlu_full_ny_professional_accounting diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_professional_law.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_professional_law.yaml new file mode 100644 index 0000000000000000000000000000000000000000..d5e2c7b7c2bb6d95a5628da15188429bacb5df42 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_professional_law.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ny_template_yaml +process_docs: !function utils.process_professional_law +tag: global_mmlu_full_ny_humanities_tasks +task: global_mmlu_full_ny_professional_law diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_professional_medicine.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_professional_medicine.yaml new file mode 100644 index 0000000000000000000000000000000000000000..cacd5df735939ab579af56ac364e8c8f37f9a272 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_professional_medicine.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ny_template_yaml +process_docs: !function utils.process_professional_medicine +tag: global_mmlu_full_ny_other_tasks +task: global_mmlu_full_ny_professional_medicine diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_professional_psychology.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_professional_psychology.yaml new file mode 100644 index 0000000000000000000000000000000000000000..ffdd86d2222fc24e7c888c917a0025aeab952c0e --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_professional_psychology.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ny_template_yaml +process_docs: !function utils.process_professional_psychology +tag: global_mmlu_full_ny_social_sciences_tasks +task: global_mmlu_full_ny_professional_psychology diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_public_relations.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_public_relations.yaml new file mode 100644 index 0000000000000000000000000000000000000000..0e6b5ab89b3178ee8fa98f681df7a9005173e903 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_public_relations.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ny_template_yaml +process_docs: !function utils.process_public_relations +tag: global_mmlu_full_ny_social_sciences_tasks +task: global_mmlu_full_ny_public_relations diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_security_studies.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_security_studies.yaml new file mode 100644 index 0000000000000000000000000000000000000000..f894fdd77c764d6894740d104bc340792bcfab97 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_security_studies.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ny_template_yaml +process_docs: !function utils.process_security_studies +tag: global_mmlu_full_ny_social_sciences_tasks +task: global_mmlu_full_ny_security_studies diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_sociology.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_sociology.yaml new file mode 100644 index 0000000000000000000000000000000000000000..1d2d0cd4d9f6acf7b8fe7b18d7aa36fe8dc75f14 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_sociology.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ny_template_yaml +process_docs: !function utils.process_sociology +tag: global_mmlu_full_ny_social_sciences_tasks +task: global_mmlu_full_ny_sociology diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_us_foreign_policy.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_us_foreign_policy.yaml new file mode 100644 index 0000000000000000000000000000000000000000..a72a237de4e1c8db883d5ac37e5d9d6e61bcf522 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_us_foreign_policy.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ny_template_yaml +process_docs: !function utils.process_us_foreign_policy +tag: global_mmlu_full_ny_social_sciences_tasks +task: global_mmlu_full_ny_us_foreign_policy diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_virology.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_virology.yaml new file mode 100644 index 0000000000000000000000000000000000000000..9eeb7cf03856f104725d6dc1cebfcf0702d90dd7 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_virology.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ny_template_yaml +process_docs: !function utils.process_virology +tag: global_mmlu_full_ny_other_tasks +task: global_mmlu_full_ny_virology diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_world_religions.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_world_religions.yaml new file mode 100644 index 0000000000000000000000000000000000000000..a1c243c8305c2680a0d1eb0cadba5deb45dbfd6c --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/global_mmlu_full_ny_world_religions.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ny_template_yaml +process_docs: !function utils.process_world_religions +tag: global_mmlu_full_ny_humanities_tasks +task: global_mmlu_full_ny_world_religions diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/utils.py b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/utils.py new file mode 100644 index 0000000000000000000000000000000000000000..7df72cb061f0fecba46e15ff9b57552817979afb --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ny/utils.py @@ -0,0 +1,73 @@ +from functools import partial + + +SUBJECTS = [ + "abstract_algebra", + "anatomy", + "astronomy", + "business_ethics", + "clinical_knowledge", + "college_biology", + "college_chemistry", + "college_computer_science", + "college_mathematics", + "college_medicine", + "college_physics", + "computer_security", + "conceptual_physics", + "econometrics", + "electrical_engineering", + "elementary_mathematics", + "formal_logic", + "global_facts", + "high_school_biology", + "high_school_chemistry", + "high_school_computer_science", + "high_school_european_history", + "high_school_geography", + "high_school_government_and_politics", + "high_school_macroeconomics", + "high_school_mathematics", + "high_school_microeconomics", + "high_school_physics", + "high_school_psychology", + "high_school_statistics", + "high_school_us_history", + "high_school_world_history", + "human_aging", + "human_sexuality", + "international_law", + "jurisprudence", + "logical_fallacies", + "machine_learning", + "management", + "marketing", + "medical_genetics", + "miscellaneous", + "moral_disputes", + "moral_scenarios", + "nutrition", + "philosophy", + "prehistory", + "professional_accounting", + "professional_law", + "professional_medicine", + "professional_psychology", + "public_relations", + "security_studies", + "sociology", + "us_foreign_policy", + "virology", + "world_religions", +] + + +def process_docs(dataset, subject): + return dataset.filter(lambda x: x["subject"] == subject) + + +process_functions = { + f"process_{subject}": partial(process_docs, subject=subject) for subject in SUBJECTS +} + +globals().update(process_functions) diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/_global_mmlu_full_pl.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/_global_mmlu_full_pl.yaml new file mode 100644 index 0000000000000000000000000000000000000000..deb51e119b8d12f1ad7e1c9355540c00428306eb --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/_global_mmlu_full_pl.yaml @@ -0,0 +1,11 @@ +group: global_mmlu_full_pl +task: + - global_mmlu_full_pl_stem + - global_mmlu_full_pl_other + - global_mmlu_full_pl_social_sciences + - global_mmlu_full_pl_humanities +aggregate_metric_list: + - metric: acc + weight_by_size: True +metadata: + version: 0.0 diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/_global_mmlu_full_pl_humanities.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/_global_mmlu_full_pl_humanities.yaml new file mode 100644 index 0000000000000000000000000000000000000000..4b5f7aa4584b551d6d6da59020c38422490c0b35 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/_global_mmlu_full_pl_humanities.yaml @@ -0,0 +1,8 @@ +group: global_mmlu_full_pl_humanities +task: + - global_mmlu_full_pl_humanities_tasks +aggregate_metric_list: + - metric: acc + weight_by_size: true +metadata: + version: 0.0 diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/_global_mmlu_full_pl_other.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/_global_mmlu_full_pl_other.yaml new file mode 100644 index 0000000000000000000000000000000000000000..241dbc1cecdd7092d7f372559beb34e6e7019d1b --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/_global_mmlu_full_pl_other.yaml @@ -0,0 +1,8 @@ +group: global_mmlu_full_pl_other +task: + - global_mmlu_full_pl_other_tasks +aggregate_metric_list: + - metric: acc + weight_by_size: true +metadata: + version: 0.0 diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/_global_mmlu_full_pl_social_sciences.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/_global_mmlu_full_pl_social_sciences.yaml new file mode 100644 index 0000000000000000000000000000000000000000..9a50a3152c08a75ba85d72b4b0b41e28500b1f0b --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/_global_mmlu_full_pl_social_sciences.yaml @@ -0,0 +1,8 @@ +group: global_mmlu_full_pl_social_sciences +task: + - global_mmlu_full_pl_social_sciences_tasks +aggregate_metric_list: + - metric: acc + weight_by_size: true +metadata: + version: 0.0 diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/_global_mmlu_full_pl_stem.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/_global_mmlu_full_pl_stem.yaml new file mode 100644 index 0000000000000000000000000000000000000000..3d11c89f119e921cc4e2301edc46f489c88afa38 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/_global_mmlu_full_pl_stem.yaml @@ -0,0 +1,8 @@ +group: global_mmlu_full_pl_stem +task: + - global_mmlu_full_pl_stem_tasks +aggregate_metric_list: + - metric: acc + weight_by_size: true +metadata: + version: 0.0 diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/_pl_template_yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/_pl_template_yaml new file mode 100644 index 0000000000000000000000000000000000000000..af8809dc2792042ed3144271bbd88c89acc8211b --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/_pl_template_yaml @@ -0,0 +1,16 @@ +dataset_path: CohereForAI/Global-MMLU +dataset_name: pl +test_split: test +fewshot_split: dev +fewshot_config: + sampler: first_n +output_type: multiple_choice +doc_to_text: "{{question.strip()}}\nA. {{option_a}}\nB. {{option_b}}\nC. {{option_c}}\nD. {{option_d}}\nAnswer:" +doc_to_choice: ["A", "B", "C", "D"] +doc_to_target: answer +metric_list: + - metric: acc + aggregation: mean + higher_is_better: true +metadata: + version: 0.0 diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_abstract_algebra.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_abstract_algebra.yaml new file mode 100644 index 0000000000000000000000000000000000000000..37f611a1b7804755d15290516e3321155dcc655c --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_abstract_algebra.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pl_template_yaml +process_docs: !function utils.process_abstract_algebra +tag: global_mmlu_full_pl_stem_tasks +task: global_mmlu_full_pl_abstract_algebra diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_anatomy.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_anatomy.yaml new file mode 100644 index 0000000000000000000000000000000000000000..c274bce178e9f12885854dedc2246993e9d44156 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_anatomy.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pl_template_yaml +process_docs: !function utils.process_anatomy +tag: global_mmlu_full_pl_stem_tasks +task: global_mmlu_full_pl_anatomy diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_astronomy.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_astronomy.yaml new file mode 100644 index 0000000000000000000000000000000000000000..99220f0ddd1d36dc346d8b778ddc333fd1f36f0d --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_astronomy.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pl_template_yaml +process_docs: !function utils.process_astronomy +tag: global_mmlu_full_pl_stem_tasks +task: global_mmlu_full_pl_astronomy diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_business_ethics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_business_ethics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..105926680ba0e802b515c0a5dcc096bffa4b51cb --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_business_ethics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pl_template_yaml +process_docs: !function utils.process_business_ethics +tag: global_mmlu_full_pl_other_tasks +task: global_mmlu_full_pl_business_ethics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_clinical_knowledge.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_clinical_knowledge.yaml new file mode 100644 index 0000000000000000000000000000000000000000..29a4fadc08d8af9e9471153fd98e3abef56fb5f7 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_clinical_knowledge.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pl_template_yaml +process_docs: !function utils.process_clinical_knowledge +tag: global_mmlu_full_pl_other_tasks +task: global_mmlu_full_pl_clinical_knowledge diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_college_biology.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_college_biology.yaml new file mode 100644 index 0000000000000000000000000000000000000000..cce1671cedb3a4163e6549ff2076f6493688a0f4 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_college_biology.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pl_template_yaml +process_docs: !function utils.process_college_biology +tag: global_mmlu_full_pl_stem_tasks +task: global_mmlu_full_pl_college_biology diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_college_chemistry.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_college_chemistry.yaml new file mode 100644 index 0000000000000000000000000000000000000000..79c6353082037914354e4ba4544d2960691d813d --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_college_chemistry.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pl_template_yaml +process_docs: !function utils.process_college_chemistry +tag: global_mmlu_full_pl_stem_tasks +task: global_mmlu_full_pl_college_chemistry diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_college_computer_science.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_college_computer_science.yaml new file mode 100644 index 0000000000000000000000000000000000000000..bb630140eb6a8601990708831e3ab712f35ea7dc --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_college_computer_science.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pl_template_yaml +process_docs: !function utils.process_college_computer_science +tag: global_mmlu_full_pl_stem_tasks +task: global_mmlu_full_pl_college_computer_science diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_college_mathematics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_college_mathematics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..6b42f767eae21b00657ec6c1a6823b9c9e8a2677 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_college_mathematics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pl_template_yaml +process_docs: !function utils.process_college_mathematics +tag: global_mmlu_full_pl_stem_tasks +task: global_mmlu_full_pl_college_mathematics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_college_medicine.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_college_medicine.yaml new file mode 100644 index 0000000000000000000000000000000000000000..43bea9764df9264452f37bb547319b1bc7ebc918 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_college_medicine.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pl_template_yaml +process_docs: !function utils.process_college_medicine +tag: global_mmlu_full_pl_other_tasks +task: global_mmlu_full_pl_college_medicine diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_college_physics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_college_physics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..0c9ea60146bd95eae494eecd34ee5f0c72f9ad47 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_college_physics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pl_template_yaml +process_docs: !function utils.process_college_physics +tag: global_mmlu_full_pl_stem_tasks +task: global_mmlu_full_pl_college_physics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_computer_security.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_computer_security.yaml new file mode 100644 index 0000000000000000000000000000000000000000..365b60a3e397e58268dd400817ba860da723fdee --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_computer_security.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pl_template_yaml +process_docs: !function utils.process_computer_security +tag: global_mmlu_full_pl_stem_tasks +task: global_mmlu_full_pl_computer_security diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_conceptual_physics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_conceptual_physics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..2b9437e3fac8256e78d438cde65a037c9d9bfbf7 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_conceptual_physics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pl_template_yaml +process_docs: !function utils.process_conceptual_physics +tag: global_mmlu_full_pl_stem_tasks +task: global_mmlu_full_pl_conceptual_physics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_econometrics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_econometrics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..648f24c422edd8433c120f37bea33493020ff8e3 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_econometrics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pl_template_yaml +process_docs: !function utils.process_econometrics +tag: global_mmlu_full_pl_social_sciences_tasks +task: global_mmlu_full_pl_econometrics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_electrical_engineering.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_electrical_engineering.yaml new file mode 100644 index 0000000000000000000000000000000000000000..196de258dee0d6d30f75f5d06575bf45418382b9 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_electrical_engineering.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pl_template_yaml +process_docs: !function utils.process_electrical_engineering +tag: global_mmlu_full_pl_stem_tasks +task: global_mmlu_full_pl_electrical_engineering diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_elementary_mathematics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_elementary_mathematics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..8646b6a393c825b2612389ebc55d102694a1f953 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_elementary_mathematics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pl_template_yaml +process_docs: !function utils.process_elementary_mathematics +tag: global_mmlu_full_pl_stem_tasks +task: global_mmlu_full_pl_elementary_mathematics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_formal_logic.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_formal_logic.yaml new file mode 100644 index 0000000000000000000000000000000000000000..2d13d283c5f5556ceebbe8ea9e0e396a6155d93a --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_formal_logic.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pl_template_yaml +process_docs: !function utils.process_formal_logic +tag: global_mmlu_full_pl_humanities_tasks +task: global_mmlu_full_pl_formal_logic diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_global_facts.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_global_facts.yaml new file mode 100644 index 0000000000000000000000000000000000000000..15bb640b93e0c3d28938631496197aa10c879161 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_global_facts.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pl_template_yaml +process_docs: !function utils.process_global_facts +tag: global_mmlu_full_pl_other_tasks +task: global_mmlu_full_pl_global_facts diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_high_school_biology.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_high_school_biology.yaml new file mode 100644 index 0000000000000000000000000000000000000000..ba96402838e5c30df22e46d46668f63633aab4e1 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_high_school_biology.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pl_template_yaml +process_docs: !function utils.process_high_school_biology +tag: global_mmlu_full_pl_stem_tasks +task: global_mmlu_full_pl_high_school_biology diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_high_school_chemistry.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_high_school_chemistry.yaml new file mode 100644 index 0000000000000000000000000000000000000000..7f142dd82f39e5d474d9bd1fa8d647fd1e2201f1 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_high_school_chemistry.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pl_template_yaml +process_docs: !function utils.process_high_school_chemistry +tag: global_mmlu_full_pl_stem_tasks +task: global_mmlu_full_pl_high_school_chemistry diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_high_school_computer_science.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_high_school_computer_science.yaml new file mode 100644 index 0000000000000000000000000000000000000000..99b3b9da11cc8bc221d2a60b10bdb89827e8c874 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_high_school_computer_science.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pl_template_yaml +process_docs: !function utils.process_high_school_computer_science +tag: global_mmlu_full_pl_stem_tasks +task: global_mmlu_full_pl_high_school_computer_science diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_high_school_european_history.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_high_school_european_history.yaml new file mode 100644 index 0000000000000000000000000000000000000000..e99b2fb93d648bcdd1977927e6eba407dfaf554c --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_high_school_european_history.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pl_template_yaml +process_docs: !function utils.process_high_school_european_history +tag: global_mmlu_full_pl_humanities_tasks +task: global_mmlu_full_pl_high_school_european_history diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_high_school_geography.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_high_school_geography.yaml new file mode 100644 index 0000000000000000000000000000000000000000..bc6113f78f15d8d51b79bafa64ea3b0a42154c41 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_high_school_geography.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pl_template_yaml +process_docs: !function utils.process_high_school_geography +tag: global_mmlu_full_pl_social_sciences_tasks +task: global_mmlu_full_pl_high_school_geography diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_high_school_government_and_politics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_high_school_government_and_politics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..05a7de9bfe7e12ea3a0eccb781e53b256de93d4e --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_high_school_government_and_politics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pl_template_yaml +process_docs: !function utils.process_high_school_government_and_politics +tag: global_mmlu_full_pl_social_sciences_tasks +task: global_mmlu_full_pl_high_school_government_and_politics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_high_school_macroeconomics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_high_school_macroeconomics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..aceda633f9b60feae440661fcbd31ac92f5bd490 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_high_school_macroeconomics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pl_template_yaml +process_docs: !function utils.process_high_school_macroeconomics +tag: global_mmlu_full_pl_social_sciences_tasks +task: global_mmlu_full_pl_high_school_macroeconomics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_high_school_mathematics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_high_school_mathematics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..6eef2cd8578fdcf1ac0d8e12584b576adafab1ee --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_high_school_mathematics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pl_template_yaml +process_docs: !function utils.process_high_school_mathematics +tag: global_mmlu_full_pl_stem_tasks +task: global_mmlu_full_pl_high_school_mathematics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_high_school_microeconomics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_high_school_microeconomics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..5adb5fa19bc7b88102d43f3b034b39bd60f794df --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_high_school_microeconomics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pl_template_yaml +process_docs: !function utils.process_high_school_microeconomics +tag: global_mmlu_full_pl_social_sciences_tasks +task: global_mmlu_full_pl_high_school_microeconomics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_high_school_physics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_high_school_physics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..fbda792097d4738c0b6890d538dc56ae5944e6e1 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_high_school_physics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pl_template_yaml +process_docs: !function utils.process_high_school_physics +tag: global_mmlu_full_pl_stem_tasks +task: global_mmlu_full_pl_high_school_physics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_high_school_psychology.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_high_school_psychology.yaml new file mode 100644 index 0000000000000000000000000000000000000000..7eb09362a7c24dc436ca10eca7ddf603d8ae70ed --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_high_school_psychology.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pl_template_yaml +process_docs: !function utils.process_high_school_psychology +tag: global_mmlu_full_pl_social_sciences_tasks +task: global_mmlu_full_pl_high_school_psychology diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_high_school_statistics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_high_school_statistics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..b7beef5b06ee6122943e89c7585778011d3a7089 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_high_school_statistics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pl_template_yaml +process_docs: !function utils.process_high_school_statistics +tag: global_mmlu_full_pl_stem_tasks +task: global_mmlu_full_pl_high_school_statistics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_high_school_us_history.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_high_school_us_history.yaml new file mode 100644 index 0000000000000000000000000000000000000000..08f45dd9c748303c2c76ba46da529ca40ab562b1 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_high_school_us_history.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pl_template_yaml +process_docs: !function utils.process_high_school_us_history +tag: global_mmlu_full_pl_humanities_tasks +task: global_mmlu_full_pl_high_school_us_history diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_high_school_world_history.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_high_school_world_history.yaml new file mode 100644 index 0000000000000000000000000000000000000000..99664de8772599a1962da0318d6afef765876b0c --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_high_school_world_history.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pl_template_yaml +process_docs: !function utils.process_high_school_world_history +tag: global_mmlu_full_pl_humanities_tasks +task: global_mmlu_full_pl_high_school_world_history diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_human_aging.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_human_aging.yaml new file mode 100644 index 0000000000000000000000000000000000000000..d63f6f8d4af5ab8920d87c2faff92879cb639dbd --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_human_aging.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pl_template_yaml +process_docs: !function utils.process_human_aging +tag: global_mmlu_full_pl_other_tasks +task: global_mmlu_full_pl_human_aging diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_human_sexuality.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_human_sexuality.yaml new file mode 100644 index 0000000000000000000000000000000000000000..8080ca8d0ebcb5d7f503d16cffaa98633c763c03 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_human_sexuality.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pl_template_yaml +process_docs: !function utils.process_human_sexuality +tag: global_mmlu_full_pl_social_sciences_tasks +task: global_mmlu_full_pl_human_sexuality diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_international_law.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_international_law.yaml new file mode 100644 index 0000000000000000000000000000000000000000..425695c1d5c64cd8b350c944e87a40027779a6dc --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_international_law.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pl_template_yaml +process_docs: !function utils.process_international_law +tag: global_mmlu_full_pl_humanities_tasks +task: global_mmlu_full_pl_international_law diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_jurisprudence.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_jurisprudence.yaml new file mode 100644 index 0000000000000000000000000000000000000000..a6455bd71724395745a31dac3e9c5fbae81b4469 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_jurisprudence.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pl_template_yaml +process_docs: !function utils.process_jurisprudence +tag: global_mmlu_full_pl_humanities_tasks +task: global_mmlu_full_pl_jurisprudence diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_logical_fallacies.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_logical_fallacies.yaml new file mode 100644 index 0000000000000000000000000000000000000000..f1359b3a24e3ca52394fb0d524c87b3b7a80b067 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_logical_fallacies.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pl_template_yaml +process_docs: !function utils.process_logical_fallacies +tag: global_mmlu_full_pl_humanities_tasks +task: global_mmlu_full_pl_logical_fallacies diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_machine_learning.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_machine_learning.yaml new file mode 100644 index 0000000000000000000000000000000000000000..3d7bb0dcc725fbb2e830772e759c44b0bc1f21a0 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_machine_learning.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pl_template_yaml +process_docs: !function utils.process_machine_learning +tag: global_mmlu_full_pl_stem_tasks +task: global_mmlu_full_pl_machine_learning diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_management.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_management.yaml new file mode 100644 index 0000000000000000000000000000000000000000..f695226c1bdcc2a798f8595a06c12c529cfa269b --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_management.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pl_template_yaml +process_docs: !function utils.process_management +tag: global_mmlu_full_pl_other_tasks +task: global_mmlu_full_pl_management diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_marketing.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_marketing.yaml new file mode 100644 index 0000000000000000000000000000000000000000..7fedcd3f8817be8bf7be01c9d3157cb18200cb77 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_marketing.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pl_template_yaml +process_docs: !function utils.process_marketing +tag: global_mmlu_full_pl_other_tasks +task: global_mmlu_full_pl_marketing diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_medical_genetics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_medical_genetics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..89da9f67565089e10cb852df563aa289afa0d455 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_medical_genetics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pl_template_yaml +process_docs: !function utils.process_medical_genetics +tag: global_mmlu_full_pl_other_tasks +task: global_mmlu_full_pl_medical_genetics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_miscellaneous.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_miscellaneous.yaml new file mode 100644 index 0000000000000000000000000000000000000000..6f34762c99cb9bbc6d273c016f4e74c91baa938b --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_miscellaneous.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pl_template_yaml +process_docs: !function utils.process_miscellaneous +tag: global_mmlu_full_pl_other_tasks +task: global_mmlu_full_pl_miscellaneous diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_moral_disputes.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_moral_disputes.yaml new file mode 100644 index 0000000000000000000000000000000000000000..25f201f4b6f16809a13a627027b1a0acfac21d46 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_moral_disputes.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pl_template_yaml +process_docs: !function utils.process_moral_disputes +tag: global_mmlu_full_pl_humanities_tasks +task: global_mmlu_full_pl_moral_disputes diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_moral_scenarios.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_moral_scenarios.yaml new file mode 100644 index 0000000000000000000000000000000000000000..fd08e6e13257a3a14a50e49c52212baca4c1f19d --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_moral_scenarios.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pl_template_yaml +process_docs: !function utils.process_moral_scenarios +tag: global_mmlu_full_pl_humanities_tasks +task: global_mmlu_full_pl_moral_scenarios diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_nutrition.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_nutrition.yaml new file mode 100644 index 0000000000000000000000000000000000000000..b61f1f17fccd40d430c7292e3337862cba73b76c --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_nutrition.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pl_template_yaml +process_docs: !function utils.process_nutrition +tag: global_mmlu_full_pl_other_tasks +task: global_mmlu_full_pl_nutrition diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_philosophy.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_philosophy.yaml new file mode 100644 index 0000000000000000000000000000000000000000..8c1bf6dc13d80759fd4aedbc131bd5a1f7777989 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_philosophy.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pl_template_yaml +process_docs: !function utils.process_philosophy +tag: global_mmlu_full_pl_humanities_tasks +task: global_mmlu_full_pl_philosophy diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_prehistory.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_prehistory.yaml new file mode 100644 index 0000000000000000000000000000000000000000..e5329e1330ba57dfdc3799bf5459eca6b00484d5 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_prehistory.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pl_template_yaml +process_docs: !function utils.process_prehistory +tag: global_mmlu_full_pl_humanities_tasks +task: global_mmlu_full_pl_prehistory diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_professional_accounting.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_professional_accounting.yaml new file mode 100644 index 0000000000000000000000000000000000000000..514b04cd1167e1b455d256bd4bd0aee450a53ac9 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_professional_accounting.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pl_template_yaml +process_docs: !function utils.process_professional_accounting +tag: global_mmlu_full_pl_other_tasks +task: global_mmlu_full_pl_professional_accounting diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_professional_law.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_professional_law.yaml new file mode 100644 index 0000000000000000000000000000000000000000..99c719f83c36178ac6aa873284d2c24742d6cc66 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_professional_law.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pl_template_yaml +process_docs: !function utils.process_professional_law +tag: global_mmlu_full_pl_humanities_tasks +task: global_mmlu_full_pl_professional_law diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_professional_medicine.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_professional_medicine.yaml new file mode 100644 index 0000000000000000000000000000000000000000..1dfafb254e809dc16186176bb61d07cccce71b45 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_professional_medicine.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pl_template_yaml +process_docs: !function utils.process_professional_medicine +tag: global_mmlu_full_pl_other_tasks +task: global_mmlu_full_pl_professional_medicine diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_professional_psychology.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_professional_psychology.yaml new file mode 100644 index 0000000000000000000000000000000000000000..5b6181c0c91f62beeba5d6f9ef2aa561438a1f0b --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_professional_psychology.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pl_template_yaml +process_docs: !function utils.process_professional_psychology +tag: global_mmlu_full_pl_social_sciences_tasks +task: global_mmlu_full_pl_professional_psychology diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_public_relations.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_public_relations.yaml new file mode 100644 index 0000000000000000000000000000000000000000..acf874dbf9630ff7b6c8118696cebbea93ea970b --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_public_relations.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pl_template_yaml +process_docs: !function utils.process_public_relations +tag: global_mmlu_full_pl_social_sciences_tasks +task: global_mmlu_full_pl_public_relations diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_security_studies.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_security_studies.yaml new file mode 100644 index 0000000000000000000000000000000000000000..d754904c4f04632e5a869e4a3478970ff9e0879a --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_security_studies.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pl_template_yaml +process_docs: !function utils.process_security_studies +tag: global_mmlu_full_pl_social_sciences_tasks +task: global_mmlu_full_pl_security_studies diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_sociology.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_sociology.yaml new file mode 100644 index 0000000000000000000000000000000000000000..4bc0fd8ff08ed950cfb842f368a0a9d8c7707b19 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_sociology.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pl_template_yaml +process_docs: !function utils.process_sociology +tag: global_mmlu_full_pl_social_sciences_tasks +task: global_mmlu_full_pl_sociology diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_us_foreign_policy.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_us_foreign_policy.yaml new file mode 100644 index 0000000000000000000000000000000000000000..ef719be1f48aab6516e2de1462b411d0eda8dddb --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_us_foreign_policy.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pl_template_yaml +process_docs: !function utils.process_us_foreign_policy +tag: global_mmlu_full_pl_social_sciences_tasks +task: global_mmlu_full_pl_us_foreign_policy diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_virology.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_virology.yaml new file mode 100644 index 0000000000000000000000000000000000000000..f9084c13da02ba1c7cf3a203404497bbcddfb3c8 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_virology.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pl_template_yaml +process_docs: !function utils.process_virology +tag: global_mmlu_full_pl_other_tasks +task: global_mmlu_full_pl_virology diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_world_religions.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_world_religions.yaml new file mode 100644 index 0000000000000000000000000000000000000000..036d0f4c0a544f0abdbc7bf5f2611d7afb20cda9 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/global_mmlu_full_pl_world_religions.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pl_template_yaml +process_docs: !function utils.process_world_religions +tag: global_mmlu_full_pl_humanities_tasks +task: global_mmlu_full_pl_world_religions diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/utils.py b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/utils.py new file mode 100644 index 0000000000000000000000000000000000000000..7df72cb061f0fecba46e15ff9b57552817979afb --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pl/utils.py @@ -0,0 +1,73 @@ +from functools import partial + + +SUBJECTS = [ + "abstract_algebra", + "anatomy", + "astronomy", + "business_ethics", + "clinical_knowledge", + "college_biology", + "college_chemistry", + "college_computer_science", + "college_mathematics", + "college_medicine", + "college_physics", + "computer_security", + "conceptual_physics", + "econometrics", + "electrical_engineering", + "elementary_mathematics", + "formal_logic", + "global_facts", + "high_school_biology", + "high_school_chemistry", + "high_school_computer_science", + "high_school_european_history", + "high_school_geography", + "high_school_government_and_politics", + "high_school_macroeconomics", + "high_school_mathematics", + "high_school_microeconomics", + "high_school_physics", + "high_school_psychology", + "high_school_statistics", + "high_school_us_history", + "high_school_world_history", + "human_aging", + "human_sexuality", + "international_law", + "jurisprudence", + "logical_fallacies", + "machine_learning", + "management", + "marketing", + "medical_genetics", + "miscellaneous", + "moral_disputes", + "moral_scenarios", + "nutrition", + "philosophy", + "prehistory", + "professional_accounting", + "professional_law", + "professional_medicine", + "professional_psychology", + "public_relations", + "security_studies", + "sociology", + "us_foreign_policy", + "virology", + "world_religions", +] + + +def process_docs(dataset, subject): + return dataset.filter(lambda x: x["subject"] == subject) + + +process_functions = { + f"process_{subject}": partial(process_docs, subject=subject) for subject in SUBJECTS +} + +globals().update(process_functions) diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/_global_mmlu_full_pt.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/_global_mmlu_full_pt.yaml new file mode 100644 index 0000000000000000000000000000000000000000..7cb9f0fe3778e4870b33733795e0e2212a282f68 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/_global_mmlu_full_pt.yaml @@ -0,0 +1,11 @@ +group: global_mmlu_full_pt +task: + - global_mmlu_full_pt_stem + - global_mmlu_full_pt_other + - global_mmlu_full_pt_social_sciences + - global_mmlu_full_pt_humanities +aggregate_metric_list: + - metric: acc + weight_by_size: True +metadata: + version: 0.0 diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/_global_mmlu_full_pt_humanities.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/_global_mmlu_full_pt_humanities.yaml new file mode 100644 index 0000000000000000000000000000000000000000..261a70288e94b635e3f41e5a4b6b2ca58d971368 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/_global_mmlu_full_pt_humanities.yaml @@ -0,0 +1,8 @@ +group: global_mmlu_full_pt_humanities +task: + - global_mmlu_full_pt_humanities_tasks +aggregate_metric_list: + - metric: acc + weight_by_size: true +metadata: + version: 0.0 diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/_global_mmlu_full_pt_other.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/_global_mmlu_full_pt_other.yaml new file mode 100644 index 0000000000000000000000000000000000000000..a61b12f504501c34dcfa4a67eff3dedc5e88ec4e --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/_global_mmlu_full_pt_other.yaml @@ -0,0 +1,8 @@ +group: global_mmlu_full_pt_other +task: + - global_mmlu_full_pt_other_tasks +aggregate_metric_list: + - metric: acc + weight_by_size: true +metadata: + version: 0.0 diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/_global_mmlu_full_pt_social_sciences.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/_global_mmlu_full_pt_social_sciences.yaml new file mode 100644 index 0000000000000000000000000000000000000000..2c04bf5a4da9fa65ffebca5c3d042cd0412e841b --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/_global_mmlu_full_pt_social_sciences.yaml @@ -0,0 +1,8 @@ +group: global_mmlu_full_pt_social_sciences +task: + - global_mmlu_full_pt_social_sciences_tasks +aggregate_metric_list: + - metric: acc + weight_by_size: true +metadata: + version: 0.0 diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/_global_mmlu_full_pt_stem.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/_global_mmlu_full_pt_stem.yaml new file mode 100644 index 0000000000000000000000000000000000000000..dc3d36104348d3a60a420360c5c47df11d108c7b --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/_global_mmlu_full_pt_stem.yaml @@ -0,0 +1,8 @@ +group: global_mmlu_full_pt_stem +task: + - global_mmlu_full_pt_stem_tasks +aggregate_metric_list: + - metric: acc + weight_by_size: true +metadata: + version: 0.0 diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/_pt_template_yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/_pt_template_yaml new file mode 100644 index 0000000000000000000000000000000000000000..66ba2417d166085f8b44cc257231a1e548166528 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/_pt_template_yaml @@ -0,0 +1,16 @@ +dataset_path: CohereForAI/Global-MMLU +dataset_name: pt +test_split: test +fewshot_split: dev +fewshot_config: + sampler: first_n +output_type: multiple_choice +doc_to_text: "{{question.strip()}}\nA. {{option_a}}\nB. {{option_b}}\nC. {{option_c}}\nD. {{option_d}}\nAnswer:" +doc_to_choice: ["A", "B", "C", "D"] +doc_to_target: answer +metric_list: + - metric: acc + aggregation: mean + higher_is_better: true +metadata: + version: 0.0 diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_abstract_algebra.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_abstract_algebra.yaml new file mode 100644 index 0000000000000000000000000000000000000000..d9efd8170b1db9b7348ce46bc50033e0b5544805 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_abstract_algebra.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pt_template_yaml +process_docs: !function utils.process_abstract_algebra +tag: global_mmlu_full_pt_stem_tasks +task: global_mmlu_full_pt_abstract_algebra diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_anatomy.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_anatomy.yaml new file mode 100644 index 0000000000000000000000000000000000000000..45390503d17e8403871c498f4e33e852aa3019eb --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_anatomy.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pt_template_yaml +process_docs: !function utils.process_anatomy +tag: global_mmlu_full_pt_stem_tasks +task: global_mmlu_full_pt_anatomy diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_astronomy.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_astronomy.yaml new file mode 100644 index 0000000000000000000000000000000000000000..90880cd077e253cafe68709633b533a02a8b0b51 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_astronomy.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pt_template_yaml +process_docs: !function utils.process_astronomy +tag: global_mmlu_full_pt_stem_tasks +task: global_mmlu_full_pt_astronomy diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_business_ethics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_business_ethics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..f18ef2d8549b1ddcc64a0b3a9c18d94eb1116ba7 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_business_ethics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pt_template_yaml +process_docs: !function utils.process_business_ethics +tag: global_mmlu_full_pt_other_tasks +task: global_mmlu_full_pt_business_ethics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_clinical_knowledge.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_clinical_knowledge.yaml new file mode 100644 index 0000000000000000000000000000000000000000..2999a02a756fa463d199050fdbd1cd579f747dfd --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_clinical_knowledge.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pt_template_yaml +process_docs: !function utils.process_clinical_knowledge +tag: global_mmlu_full_pt_other_tasks +task: global_mmlu_full_pt_clinical_knowledge diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_college_biology.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_college_biology.yaml new file mode 100644 index 0000000000000000000000000000000000000000..0cf0a61b55d9c7283755b4d7c1f61c66bbf2a95a --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_college_biology.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pt_template_yaml +process_docs: !function utils.process_college_biology +tag: global_mmlu_full_pt_stem_tasks +task: global_mmlu_full_pt_college_biology diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_college_chemistry.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_college_chemistry.yaml new file mode 100644 index 0000000000000000000000000000000000000000..91d8cd2ed1e520094c8b3253cfd17df2c6f1dfc7 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_college_chemistry.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pt_template_yaml +process_docs: !function utils.process_college_chemistry +tag: global_mmlu_full_pt_stem_tasks +task: global_mmlu_full_pt_college_chemistry diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_college_computer_science.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_college_computer_science.yaml new file mode 100644 index 0000000000000000000000000000000000000000..68592aafe5b5b00c38bad52edde00f8535a32fde --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_college_computer_science.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pt_template_yaml +process_docs: !function utils.process_college_computer_science +tag: global_mmlu_full_pt_stem_tasks +task: global_mmlu_full_pt_college_computer_science diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_college_mathematics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_college_mathematics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..31d7f6afda5cfafd2531f8eda65a13c676f498de --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_college_mathematics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pt_template_yaml +process_docs: !function utils.process_college_mathematics +tag: global_mmlu_full_pt_stem_tasks +task: global_mmlu_full_pt_college_mathematics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_college_medicine.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_college_medicine.yaml new file mode 100644 index 0000000000000000000000000000000000000000..46ec823241d6af4e85b37df2d4649260d4de5c66 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_college_medicine.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pt_template_yaml +process_docs: !function utils.process_college_medicine +tag: global_mmlu_full_pt_other_tasks +task: global_mmlu_full_pt_college_medicine diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_college_physics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_college_physics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..2cf6402dd6ec95592e34932a35ef8b67829a61d3 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_college_physics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pt_template_yaml +process_docs: !function utils.process_college_physics +tag: global_mmlu_full_pt_stem_tasks +task: global_mmlu_full_pt_college_physics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_computer_security.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_computer_security.yaml new file mode 100644 index 0000000000000000000000000000000000000000..0953a105a72a03a2f2109ad7bb47eba29995c26a --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_computer_security.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pt_template_yaml +process_docs: !function utils.process_computer_security +tag: global_mmlu_full_pt_stem_tasks +task: global_mmlu_full_pt_computer_security diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_conceptual_physics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_conceptual_physics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..0e6e91a9935da462b6f5bc101fecc2216f3a68b5 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_conceptual_physics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pt_template_yaml +process_docs: !function utils.process_conceptual_physics +tag: global_mmlu_full_pt_stem_tasks +task: global_mmlu_full_pt_conceptual_physics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_econometrics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_econometrics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..67c29915c614a1d01661a72df1a56e27f6a54f6e --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_econometrics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pt_template_yaml +process_docs: !function utils.process_econometrics +tag: global_mmlu_full_pt_social_sciences_tasks +task: global_mmlu_full_pt_econometrics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_electrical_engineering.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_electrical_engineering.yaml new file mode 100644 index 0000000000000000000000000000000000000000..5a6ba82e29b272d00fffd2374659d010ae8b0f9a --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_electrical_engineering.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pt_template_yaml +process_docs: !function utils.process_electrical_engineering +tag: global_mmlu_full_pt_stem_tasks +task: global_mmlu_full_pt_electrical_engineering diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_elementary_mathematics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_elementary_mathematics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..3d66a664c39f1b46ab6e77ae665e57236c334e75 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_elementary_mathematics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pt_template_yaml +process_docs: !function utils.process_elementary_mathematics +tag: global_mmlu_full_pt_stem_tasks +task: global_mmlu_full_pt_elementary_mathematics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_formal_logic.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_formal_logic.yaml new file mode 100644 index 0000000000000000000000000000000000000000..683d6dddcd054ffc1219e6db310714d70c74b04e --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_formal_logic.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pt_template_yaml +process_docs: !function utils.process_formal_logic +tag: global_mmlu_full_pt_humanities_tasks +task: global_mmlu_full_pt_formal_logic diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_global_facts.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_global_facts.yaml new file mode 100644 index 0000000000000000000000000000000000000000..e4396542a55f7f433bcc361a7dd27f3ef21e365e --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_global_facts.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pt_template_yaml +process_docs: !function utils.process_global_facts +tag: global_mmlu_full_pt_other_tasks +task: global_mmlu_full_pt_global_facts diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_high_school_biology.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_high_school_biology.yaml new file mode 100644 index 0000000000000000000000000000000000000000..89fefd1c32a73e2beb44d383632ea1e50b8ec63f --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_high_school_biology.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pt_template_yaml +process_docs: !function utils.process_high_school_biology +tag: global_mmlu_full_pt_stem_tasks +task: global_mmlu_full_pt_high_school_biology diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_high_school_chemistry.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_high_school_chemistry.yaml new file mode 100644 index 0000000000000000000000000000000000000000..ea323d8a64f0701fb95f14b306a175af2227bb13 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_high_school_chemistry.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pt_template_yaml +process_docs: !function utils.process_high_school_chemistry +tag: global_mmlu_full_pt_stem_tasks +task: global_mmlu_full_pt_high_school_chemistry diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_high_school_computer_science.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_high_school_computer_science.yaml new file mode 100644 index 0000000000000000000000000000000000000000..5f8f008235f7fdb1a851babb6935645144d24fcd --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_high_school_computer_science.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pt_template_yaml +process_docs: !function utils.process_high_school_computer_science +tag: global_mmlu_full_pt_stem_tasks +task: global_mmlu_full_pt_high_school_computer_science diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_high_school_european_history.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_high_school_european_history.yaml new file mode 100644 index 0000000000000000000000000000000000000000..bef7a31683e7bd64839b1e67a47ab395fecd2507 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_high_school_european_history.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pt_template_yaml +process_docs: !function utils.process_high_school_european_history +tag: global_mmlu_full_pt_humanities_tasks +task: global_mmlu_full_pt_high_school_european_history diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_high_school_geography.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_high_school_geography.yaml new file mode 100644 index 0000000000000000000000000000000000000000..e69c297805e5137c32371408199f7aed427ce273 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_high_school_geography.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pt_template_yaml +process_docs: !function utils.process_high_school_geography +tag: global_mmlu_full_pt_social_sciences_tasks +task: global_mmlu_full_pt_high_school_geography diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_high_school_government_and_politics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_high_school_government_and_politics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..e3fa920d7246a6e88c656bc3bce9270809f9d7ea --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_high_school_government_and_politics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pt_template_yaml +process_docs: !function utils.process_high_school_government_and_politics +tag: global_mmlu_full_pt_social_sciences_tasks +task: global_mmlu_full_pt_high_school_government_and_politics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_high_school_macroeconomics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_high_school_macroeconomics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..6b7ca2f4e826b4b3116b7d4f21e8f4de04e73d33 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_high_school_macroeconomics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pt_template_yaml +process_docs: !function utils.process_high_school_macroeconomics +tag: global_mmlu_full_pt_social_sciences_tasks +task: global_mmlu_full_pt_high_school_macroeconomics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_high_school_mathematics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_high_school_mathematics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..4713674dcbb43d0460556eb63e91c0ffd1bed583 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_high_school_mathematics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pt_template_yaml +process_docs: !function utils.process_high_school_mathematics +tag: global_mmlu_full_pt_stem_tasks +task: global_mmlu_full_pt_high_school_mathematics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_high_school_microeconomics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_high_school_microeconomics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..d6475e995c6ef42e7dea9e14262ccbc8cdd6f0b8 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_high_school_microeconomics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pt_template_yaml +process_docs: !function utils.process_high_school_microeconomics +tag: global_mmlu_full_pt_social_sciences_tasks +task: global_mmlu_full_pt_high_school_microeconomics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_high_school_physics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_high_school_physics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..9eaed31adcac907953964d9cbfb46f563030a4a1 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_high_school_physics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pt_template_yaml +process_docs: !function utils.process_high_school_physics +tag: global_mmlu_full_pt_stem_tasks +task: global_mmlu_full_pt_high_school_physics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_high_school_psychology.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_high_school_psychology.yaml new file mode 100644 index 0000000000000000000000000000000000000000..d09e1eb97940084b3f35e4a212f9492c68f20143 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_high_school_psychology.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pt_template_yaml +process_docs: !function utils.process_high_school_psychology +tag: global_mmlu_full_pt_social_sciences_tasks +task: global_mmlu_full_pt_high_school_psychology diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_high_school_statistics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_high_school_statistics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..3d8c1447faca7b6a41e644ef705c0ffa7b4acd88 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_high_school_statistics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pt_template_yaml +process_docs: !function utils.process_high_school_statistics +tag: global_mmlu_full_pt_stem_tasks +task: global_mmlu_full_pt_high_school_statistics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_high_school_us_history.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_high_school_us_history.yaml new file mode 100644 index 0000000000000000000000000000000000000000..a883b438ef06e82e986f53d1e91c71757368c312 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_high_school_us_history.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pt_template_yaml +process_docs: !function utils.process_high_school_us_history +tag: global_mmlu_full_pt_humanities_tasks +task: global_mmlu_full_pt_high_school_us_history diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_high_school_world_history.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_high_school_world_history.yaml new file mode 100644 index 0000000000000000000000000000000000000000..6ea1454e35beec1ba65d7235b7d984768b45e09d --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_high_school_world_history.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pt_template_yaml +process_docs: !function utils.process_high_school_world_history +tag: global_mmlu_full_pt_humanities_tasks +task: global_mmlu_full_pt_high_school_world_history diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_human_aging.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_human_aging.yaml new file mode 100644 index 0000000000000000000000000000000000000000..34033c55b88be89f49997de2820f88eb689d2e2d --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_human_aging.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pt_template_yaml +process_docs: !function utils.process_human_aging +tag: global_mmlu_full_pt_other_tasks +task: global_mmlu_full_pt_human_aging diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_human_sexuality.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_human_sexuality.yaml new file mode 100644 index 0000000000000000000000000000000000000000..bf961c33a549b15099087ec627745a5e273f11a0 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_human_sexuality.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pt_template_yaml +process_docs: !function utils.process_human_sexuality +tag: global_mmlu_full_pt_social_sciences_tasks +task: global_mmlu_full_pt_human_sexuality diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_international_law.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_international_law.yaml new file mode 100644 index 0000000000000000000000000000000000000000..5247fc9bd030a7c43b08c3146ef05738449bf953 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_international_law.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pt_template_yaml +process_docs: !function utils.process_international_law +tag: global_mmlu_full_pt_humanities_tasks +task: global_mmlu_full_pt_international_law diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_jurisprudence.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_jurisprudence.yaml new file mode 100644 index 0000000000000000000000000000000000000000..07e78da50b5eae4d39f36a9e39560826618f950f --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_jurisprudence.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pt_template_yaml +process_docs: !function utils.process_jurisprudence +tag: global_mmlu_full_pt_humanities_tasks +task: global_mmlu_full_pt_jurisprudence diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_logical_fallacies.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_logical_fallacies.yaml new file mode 100644 index 0000000000000000000000000000000000000000..c2451399b8c66c2394729a43dbab2525b3d4d501 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_logical_fallacies.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pt_template_yaml +process_docs: !function utils.process_logical_fallacies +tag: global_mmlu_full_pt_humanities_tasks +task: global_mmlu_full_pt_logical_fallacies diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_machine_learning.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_machine_learning.yaml new file mode 100644 index 0000000000000000000000000000000000000000..79c577eac708a196c8015b8c0413f884a275b8e0 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_machine_learning.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pt_template_yaml +process_docs: !function utils.process_machine_learning +tag: global_mmlu_full_pt_stem_tasks +task: global_mmlu_full_pt_machine_learning diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_management.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_management.yaml new file mode 100644 index 0000000000000000000000000000000000000000..a344b1c32eaa834f8dd4f84d08914cf71c8ea099 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_management.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pt_template_yaml +process_docs: !function utils.process_management +tag: global_mmlu_full_pt_other_tasks +task: global_mmlu_full_pt_management diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_marketing.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_marketing.yaml new file mode 100644 index 0000000000000000000000000000000000000000..eeff36b979929dd011e94c56c78d418d9ef9da55 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_marketing.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pt_template_yaml +process_docs: !function utils.process_marketing +tag: global_mmlu_full_pt_other_tasks +task: global_mmlu_full_pt_marketing diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_medical_genetics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_medical_genetics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..27985380f5b9d2e9aed8b8c67965c0893d78c26b --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_medical_genetics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pt_template_yaml +process_docs: !function utils.process_medical_genetics +tag: global_mmlu_full_pt_other_tasks +task: global_mmlu_full_pt_medical_genetics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_miscellaneous.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_miscellaneous.yaml new file mode 100644 index 0000000000000000000000000000000000000000..e2fa1da1c1838c94b0e4a3a12a75928d0e020514 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_miscellaneous.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pt_template_yaml +process_docs: !function utils.process_miscellaneous +tag: global_mmlu_full_pt_other_tasks +task: global_mmlu_full_pt_miscellaneous diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_moral_disputes.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_moral_disputes.yaml new file mode 100644 index 0000000000000000000000000000000000000000..e83d186e2f9ee63dbce9e5559618f77443068d42 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_moral_disputes.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pt_template_yaml +process_docs: !function utils.process_moral_disputes +tag: global_mmlu_full_pt_humanities_tasks +task: global_mmlu_full_pt_moral_disputes diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_moral_scenarios.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_moral_scenarios.yaml new file mode 100644 index 0000000000000000000000000000000000000000..3529a15c3b0e6f06b2cc7c037965b0232b03111b --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_moral_scenarios.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pt_template_yaml +process_docs: !function utils.process_moral_scenarios +tag: global_mmlu_full_pt_humanities_tasks +task: global_mmlu_full_pt_moral_scenarios diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_nutrition.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_nutrition.yaml new file mode 100644 index 0000000000000000000000000000000000000000..e51eefe02a33c8252f308a6cca66b7a17365af9f --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_nutrition.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pt_template_yaml +process_docs: !function utils.process_nutrition +tag: global_mmlu_full_pt_other_tasks +task: global_mmlu_full_pt_nutrition diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_philosophy.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_philosophy.yaml new file mode 100644 index 0000000000000000000000000000000000000000..ec0826b95c63da87f3834a0f9dad7c744895e291 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_philosophy.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pt_template_yaml +process_docs: !function utils.process_philosophy +tag: global_mmlu_full_pt_humanities_tasks +task: global_mmlu_full_pt_philosophy diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_prehistory.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_prehistory.yaml new file mode 100644 index 0000000000000000000000000000000000000000..324dfe69ed57361ae56f48f16c72a431e86bc503 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_prehistory.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pt_template_yaml +process_docs: !function utils.process_prehistory +tag: global_mmlu_full_pt_humanities_tasks +task: global_mmlu_full_pt_prehistory diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_professional_accounting.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_professional_accounting.yaml new file mode 100644 index 0000000000000000000000000000000000000000..530c918efd7bf3055df23096dde71892282efafc --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_professional_accounting.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pt_template_yaml +process_docs: !function utils.process_professional_accounting +tag: global_mmlu_full_pt_other_tasks +task: global_mmlu_full_pt_professional_accounting diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_professional_law.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_professional_law.yaml new file mode 100644 index 0000000000000000000000000000000000000000..f7a3679c9845fa23a25dac7f39a1f504efdbc88b --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_professional_law.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pt_template_yaml +process_docs: !function utils.process_professional_law +tag: global_mmlu_full_pt_humanities_tasks +task: global_mmlu_full_pt_professional_law diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_professional_medicine.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_professional_medicine.yaml new file mode 100644 index 0000000000000000000000000000000000000000..0f4cc0066a1f5679d3529802fd5a9e561828e733 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_professional_medicine.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pt_template_yaml +process_docs: !function utils.process_professional_medicine +tag: global_mmlu_full_pt_other_tasks +task: global_mmlu_full_pt_professional_medicine diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_professional_psychology.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_professional_psychology.yaml new file mode 100644 index 0000000000000000000000000000000000000000..4c5884c9f668eb9816d6670260447e57bcab3dc5 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_professional_psychology.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pt_template_yaml +process_docs: !function utils.process_professional_psychology +tag: global_mmlu_full_pt_social_sciences_tasks +task: global_mmlu_full_pt_professional_psychology diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_public_relations.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_public_relations.yaml new file mode 100644 index 0000000000000000000000000000000000000000..bb2d65360ab95ef7c128afa27d2126dbe75070aa --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_public_relations.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pt_template_yaml +process_docs: !function utils.process_public_relations +tag: global_mmlu_full_pt_social_sciences_tasks +task: global_mmlu_full_pt_public_relations diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_security_studies.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_security_studies.yaml new file mode 100644 index 0000000000000000000000000000000000000000..1af8d6621226920b8319bcedde1eee2d21b45b5a --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_security_studies.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pt_template_yaml +process_docs: !function utils.process_security_studies +tag: global_mmlu_full_pt_social_sciences_tasks +task: global_mmlu_full_pt_security_studies diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_sociology.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_sociology.yaml new file mode 100644 index 0000000000000000000000000000000000000000..3ef8fcb7972b983ccdb8ca46ccb3616b0371d2dc --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_sociology.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pt_template_yaml +process_docs: !function utils.process_sociology +tag: global_mmlu_full_pt_social_sciences_tasks +task: global_mmlu_full_pt_sociology diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_us_foreign_policy.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_us_foreign_policy.yaml new file mode 100644 index 0000000000000000000000000000000000000000..8b48f52873a61574601312edce12c244fb550dfd --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_us_foreign_policy.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pt_template_yaml +process_docs: !function utils.process_us_foreign_policy +tag: global_mmlu_full_pt_social_sciences_tasks +task: global_mmlu_full_pt_us_foreign_policy diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_virology.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_virology.yaml new file mode 100644 index 0000000000000000000000000000000000000000..4b0de75322e453c0b3049c2fc42e952c3a8ec1a2 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_virology.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pt_template_yaml +process_docs: !function utils.process_virology +tag: global_mmlu_full_pt_other_tasks +task: global_mmlu_full_pt_virology diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_world_religions.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_world_religions.yaml new file mode 100644 index 0000000000000000000000000000000000000000..796485864c8e686a637f68bcecf3a32136368138 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/global_mmlu_full_pt_world_religions.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _pt_template_yaml +process_docs: !function utils.process_world_religions +tag: global_mmlu_full_pt_humanities_tasks +task: global_mmlu_full_pt_world_religions diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/utils.py b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/utils.py new file mode 100644 index 0000000000000000000000000000000000000000..7df72cb061f0fecba46e15ff9b57552817979afb --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/pt/utils.py @@ -0,0 +1,73 @@ +from functools import partial + + +SUBJECTS = [ + "abstract_algebra", + "anatomy", + "astronomy", + "business_ethics", + "clinical_knowledge", + "college_biology", + "college_chemistry", + "college_computer_science", + "college_mathematics", + "college_medicine", + "college_physics", + "computer_security", + "conceptual_physics", + "econometrics", + "electrical_engineering", + "elementary_mathematics", + "formal_logic", + "global_facts", + "high_school_biology", + "high_school_chemistry", + "high_school_computer_science", + "high_school_european_history", + "high_school_geography", + "high_school_government_and_politics", + "high_school_macroeconomics", + "high_school_mathematics", + "high_school_microeconomics", + "high_school_physics", + "high_school_psychology", + "high_school_statistics", + "high_school_us_history", + "high_school_world_history", + "human_aging", + "human_sexuality", + "international_law", + "jurisprudence", + "logical_fallacies", + "machine_learning", + "management", + "marketing", + "medical_genetics", + "miscellaneous", + "moral_disputes", + "moral_scenarios", + "nutrition", + "philosophy", + "prehistory", + "professional_accounting", + "professional_law", + "professional_medicine", + "professional_psychology", + "public_relations", + "security_studies", + "sociology", + "us_foreign_policy", + "virology", + "world_religions", +] + + +def process_docs(dataset, subject): + return dataset.filter(lambda x: x["subject"] == subject) + + +process_functions = { + f"process_{subject}": partial(process_docs, subject=subject) for subject in SUBJECTS +} + +globals().update(process_functions) diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/_global_mmlu_full_ro.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/_global_mmlu_full_ro.yaml new file mode 100644 index 0000000000000000000000000000000000000000..f4e5575f04ced9e99d0785ddf1b6ff06654af239 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/_global_mmlu_full_ro.yaml @@ -0,0 +1,11 @@ +group: global_mmlu_full_ro +task: + - global_mmlu_full_ro_stem + - global_mmlu_full_ro_other + - global_mmlu_full_ro_social_sciences + - global_mmlu_full_ro_humanities +aggregate_metric_list: + - metric: acc + weight_by_size: True +metadata: + version: 0.0 diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/_global_mmlu_full_ro_humanities.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/_global_mmlu_full_ro_humanities.yaml new file mode 100644 index 0000000000000000000000000000000000000000..d54268b0af5ffe2cbdd10d3295aff7176b6b4819 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/_global_mmlu_full_ro_humanities.yaml @@ -0,0 +1,8 @@ +group: global_mmlu_full_ro_humanities +task: + - global_mmlu_full_ro_humanities_tasks +aggregate_metric_list: + - metric: acc + weight_by_size: true +metadata: + version: 0.0 diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/_global_mmlu_full_ro_other.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/_global_mmlu_full_ro_other.yaml new file mode 100644 index 0000000000000000000000000000000000000000..4e58aea9a5d51491b7f23b4976124abec6de75b9 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/_global_mmlu_full_ro_other.yaml @@ -0,0 +1,8 @@ +group: global_mmlu_full_ro_other +task: + - global_mmlu_full_ro_other_tasks +aggregate_metric_list: + - metric: acc + weight_by_size: true +metadata: + version: 0.0 diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/_global_mmlu_full_ro_social_sciences.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/_global_mmlu_full_ro_social_sciences.yaml new file mode 100644 index 0000000000000000000000000000000000000000..e1cb84a91129f805f5391e442c84c2ae37967ebc --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/_global_mmlu_full_ro_social_sciences.yaml @@ -0,0 +1,8 @@ +group: global_mmlu_full_ro_social_sciences +task: + - global_mmlu_full_ro_social_sciences_tasks +aggregate_metric_list: + - metric: acc + weight_by_size: true +metadata: + version: 0.0 diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/_global_mmlu_full_ro_stem.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/_global_mmlu_full_ro_stem.yaml new file mode 100644 index 0000000000000000000000000000000000000000..de0e406fac5194a846713eb63cfa172003515ac0 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/_global_mmlu_full_ro_stem.yaml @@ -0,0 +1,8 @@ +group: global_mmlu_full_ro_stem +task: + - global_mmlu_full_ro_stem_tasks +aggregate_metric_list: + - metric: acc + weight_by_size: true +metadata: + version: 0.0 diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/_ro_template_yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/_ro_template_yaml new file mode 100644 index 0000000000000000000000000000000000000000..e5cb6dd0ecc8d32e03797c70cd3414351e14f63d --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/_ro_template_yaml @@ -0,0 +1,16 @@ +dataset_path: CohereForAI/Global-MMLU +dataset_name: ro +test_split: test +fewshot_split: dev +fewshot_config: + sampler: first_n +output_type: multiple_choice +doc_to_text: "{{question.strip()}}\nA. {{option_a}}\nB. {{option_b}}\nC. {{option_c}}\nD. {{option_d}}\nAnswer:" +doc_to_choice: ["A", "B", "C", "D"] +doc_to_target: answer +metric_list: + - metric: acc + aggregation: mean + higher_is_better: true +metadata: + version: 0.0 diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_abstract_algebra.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_abstract_algebra.yaml new file mode 100644 index 0000000000000000000000000000000000000000..c505fb8b204fc910e30e068462a1dd74111fffce --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_abstract_algebra.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ro_template_yaml +process_docs: !function utils.process_abstract_algebra +tag: global_mmlu_full_ro_stem_tasks +task: global_mmlu_full_ro_abstract_algebra diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_anatomy.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_anatomy.yaml new file mode 100644 index 0000000000000000000000000000000000000000..0c13018cdb68e55e3f55e7deb4e8a69cad49c749 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_anatomy.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ro_template_yaml +process_docs: !function utils.process_anatomy +tag: global_mmlu_full_ro_stem_tasks +task: global_mmlu_full_ro_anatomy diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_astronomy.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_astronomy.yaml new file mode 100644 index 0000000000000000000000000000000000000000..9f4caefba6f1651703a66f6ce51d5bd2768fe410 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_astronomy.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ro_template_yaml +process_docs: !function utils.process_astronomy +tag: global_mmlu_full_ro_stem_tasks +task: global_mmlu_full_ro_astronomy diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_business_ethics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_business_ethics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..1c1387fd182fc7aa9fb098e2b9482724728fbea3 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_business_ethics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ro_template_yaml +process_docs: !function utils.process_business_ethics +tag: global_mmlu_full_ro_other_tasks +task: global_mmlu_full_ro_business_ethics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_clinical_knowledge.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_clinical_knowledge.yaml new file mode 100644 index 0000000000000000000000000000000000000000..b9e0dbb473cd251e83ad798d2a9698eda30fb5a0 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_clinical_knowledge.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ro_template_yaml +process_docs: !function utils.process_clinical_knowledge +tag: global_mmlu_full_ro_other_tasks +task: global_mmlu_full_ro_clinical_knowledge diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_college_biology.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_college_biology.yaml new file mode 100644 index 0000000000000000000000000000000000000000..5bf14ab0f5d8a36b07a06a5b6d635bae4b2c5e28 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_college_biology.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ro_template_yaml +process_docs: !function utils.process_college_biology +tag: global_mmlu_full_ro_stem_tasks +task: global_mmlu_full_ro_college_biology diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_college_chemistry.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_college_chemistry.yaml new file mode 100644 index 0000000000000000000000000000000000000000..5903474436e9ea97e7fd86cb1e3988132e706fdb --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_college_chemistry.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ro_template_yaml +process_docs: !function utils.process_college_chemistry +tag: global_mmlu_full_ro_stem_tasks +task: global_mmlu_full_ro_college_chemistry diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_college_computer_science.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_college_computer_science.yaml new file mode 100644 index 0000000000000000000000000000000000000000..6bb64c2e201d7722ad353f1db87e38f14b0a3baf --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_college_computer_science.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ro_template_yaml +process_docs: !function utils.process_college_computer_science +tag: global_mmlu_full_ro_stem_tasks +task: global_mmlu_full_ro_college_computer_science diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_college_mathematics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_college_mathematics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..d719a5ef98b6aa7be608a6e100583c75ae874aaf --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_college_mathematics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ro_template_yaml +process_docs: !function utils.process_college_mathematics +tag: global_mmlu_full_ro_stem_tasks +task: global_mmlu_full_ro_college_mathematics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_college_medicine.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_college_medicine.yaml new file mode 100644 index 0000000000000000000000000000000000000000..c9284a8f9cf6ba9c892be6251b77432f41db9fd1 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_college_medicine.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ro_template_yaml +process_docs: !function utils.process_college_medicine +tag: global_mmlu_full_ro_other_tasks +task: global_mmlu_full_ro_college_medicine diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_college_physics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_college_physics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..1d27d843707119cae640c69659f23da770ccb3ed --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_college_physics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ro_template_yaml +process_docs: !function utils.process_college_physics +tag: global_mmlu_full_ro_stem_tasks +task: global_mmlu_full_ro_college_physics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_computer_security.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_computer_security.yaml new file mode 100644 index 0000000000000000000000000000000000000000..1d63556e36c5ccd350b1bf0a5419ff8a44554fca --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_computer_security.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ro_template_yaml +process_docs: !function utils.process_computer_security +tag: global_mmlu_full_ro_stem_tasks +task: global_mmlu_full_ro_computer_security diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_conceptual_physics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_conceptual_physics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..25f30a368d61f9a6a240db0b657a262718ab43a6 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_conceptual_physics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ro_template_yaml +process_docs: !function utils.process_conceptual_physics +tag: global_mmlu_full_ro_stem_tasks +task: global_mmlu_full_ro_conceptual_physics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_econometrics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_econometrics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..1fa6b5d23cd039f9b3e194620cd0c1dd84f5c7c9 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_econometrics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ro_template_yaml +process_docs: !function utils.process_econometrics +tag: global_mmlu_full_ro_social_sciences_tasks +task: global_mmlu_full_ro_econometrics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_electrical_engineering.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_electrical_engineering.yaml new file mode 100644 index 0000000000000000000000000000000000000000..f6eb4b6e4ca6ed74a7ec8a26afc90d63ced323e1 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_electrical_engineering.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ro_template_yaml +process_docs: !function utils.process_electrical_engineering +tag: global_mmlu_full_ro_stem_tasks +task: global_mmlu_full_ro_electrical_engineering diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_elementary_mathematics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_elementary_mathematics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..e99772e27ec8de94e5ec2773ff8147c942a37c56 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_elementary_mathematics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ro_template_yaml +process_docs: !function utils.process_elementary_mathematics +tag: global_mmlu_full_ro_stem_tasks +task: global_mmlu_full_ro_elementary_mathematics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_formal_logic.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_formal_logic.yaml new file mode 100644 index 0000000000000000000000000000000000000000..be99bd00764538247bd19da55a1c12a1b99a3344 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_formal_logic.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ro_template_yaml +process_docs: !function utils.process_formal_logic +tag: global_mmlu_full_ro_humanities_tasks +task: global_mmlu_full_ro_formal_logic diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_global_facts.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_global_facts.yaml new file mode 100644 index 0000000000000000000000000000000000000000..819937e79b6b56eb78107ae1331df479c8c73f2d --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_global_facts.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ro_template_yaml +process_docs: !function utils.process_global_facts +tag: global_mmlu_full_ro_other_tasks +task: global_mmlu_full_ro_global_facts diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_high_school_biology.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_high_school_biology.yaml new file mode 100644 index 0000000000000000000000000000000000000000..d7509581cae17e6a7800017be253b4617ddfab48 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_high_school_biology.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ro_template_yaml +process_docs: !function utils.process_high_school_biology +tag: global_mmlu_full_ro_stem_tasks +task: global_mmlu_full_ro_high_school_biology diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_high_school_chemistry.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_high_school_chemistry.yaml new file mode 100644 index 0000000000000000000000000000000000000000..d089583f6f3afdc5bd9b07a315c6cff8f42fd2cb --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_high_school_chemistry.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ro_template_yaml +process_docs: !function utils.process_high_school_chemistry +tag: global_mmlu_full_ro_stem_tasks +task: global_mmlu_full_ro_high_school_chemistry diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_high_school_computer_science.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_high_school_computer_science.yaml new file mode 100644 index 0000000000000000000000000000000000000000..46d5f4723ff084e157c09a2468ab28586bb6902d --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_high_school_computer_science.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ro_template_yaml +process_docs: !function utils.process_high_school_computer_science +tag: global_mmlu_full_ro_stem_tasks +task: global_mmlu_full_ro_high_school_computer_science diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_high_school_european_history.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_high_school_european_history.yaml new file mode 100644 index 0000000000000000000000000000000000000000..1a1ae7e72571b56dac54b7f4cac108ee35a3be9e --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_high_school_european_history.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ro_template_yaml +process_docs: !function utils.process_high_school_european_history +tag: global_mmlu_full_ro_humanities_tasks +task: global_mmlu_full_ro_high_school_european_history diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_high_school_geography.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_high_school_geography.yaml new file mode 100644 index 0000000000000000000000000000000000000000..92935be5c49c374198604ef6c3c4597c6e6178bb --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_high_school_geography.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ro_template_yaml +process_docs: !function utils.process_high_school_geography +tag: global_mmlu_full_ro_social_sciences_tasks +task: global_mmlu_full_ro_high_school_geography diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_high_school_government_and_politics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_high_school_government_and_politics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..efd2a03fa59de3e31e0e8aaed19868f9ebb6611b --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_high_school_government_and_politics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ro_template_yaml +process_docs: !function utils.process_high_school_government_and_politics +tag: global_mmlu_full_ro_social_sciences_tasks +task: global_mmlu_full_ro_high_school_government_and_politics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_high_school_macroeconomics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_high_school_macroeconomics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..fe2f97d1f167d2c37e67f85bfff4d4c5c8b9177a --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_high_school_macroeconomics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ro_template_yaml +process_docs: !function utils.process_high_school_macroeconomics +tag: global_mmlu_full_ro_social_sciences_tasks +task: global_mmlu_full_ro_high_school_macroeconomics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_high_school_mathematics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_high_school_mathematics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..f0432a01f1bb3511040ece9d4305953452d5a95d --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_high_school_mathematics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ro_template_yaml +process_docs: !function utils.process_high_school_mathematics +tag: global_mmlu_full_ro_stem_tasks +task: global_mmlu_full_ro_high_school_mathematics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_high_school_microeconomics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_high_school_microeconomics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..507fab8693c7962e681e58bcf64b76d8b0da8649 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_high_school_microeconomics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ro_template_yaml +process_docs: !function utils.process_high_school_microeconomics +tag: global_mmlu_full_ro_social_sciences_tasks +task: global_mmlu_full_ro_high_school_microeconomics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_high_school_physics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_high_school_physics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..19a76707ef5e11645afa82538a84c1a57f76c2b7 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_high_school_physics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ro_template_yaml +process_docs: !function utils.process_high_school_physics +tag: global_mmlu_full_ro_stem_tasks +task: global_mmlu_full_ro_high_school_physics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_high_school_psychology.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_high_school_psychology.yaml new file mode 100644 index 0000000000000000000000000000000000000000..d27fc262eeec0f83d51323082afd596b3e86e71f --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_high_school_psychology.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ro_template_yaml +process_docs: !function utils.process_high_school_psychology +tag: global_mmlu_full_ro_social_sciences_tasks +task: global_mmlu_full_ro_high_school_psychology diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_high_school_statistics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_high_school_statistics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..8f8023bcfc7dab66b4090240a6a73d95f117b1c9 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_high_school_statistics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ro_template_yaml +process_docs: !function utils.process_high_school_statistics +tag: global_mmlu_full_ro_stem_tasks +task: global_mmlu_full_ro_high_school_statistics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_high_school_us_history.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_high_school_us_history.yaml new file mode 100644 index 0000000000000000000000000000000000000000..acc5fc418423e59fcb7e6afe369f69e701eb6d55 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_high_school_us_history.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ro_template_yaml +process_docs: !function utils.process_high_school_us_history +tag: global_mmlu_full_ro_humanities_tasks +task: global_mmlu_full_ro_high_school_us_history diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_high_school_world_history.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_high_school_world_history.yaml new file mode 100644 index 0000000000000000000000000000000000000000..9ea7c933656e09cfe84d071299b74af234c08e6e --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_high_school_world_history.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ro_template_yaml +process_docs: !function utils.process_high_school_world_history +tag: global_mmlu_full_ro_humanities_tasks +task: global_mmlu_full_ro_high_school_world_history diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_human_aging.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_human_aging.yaml new file mode 100644 index 0000000000000000000000000000000000000000..6b984c55b4a5550373eed5879c85bbb0f2b0d099 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_human_aging.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ro_template_yaml +process_docs: !function utils.process_human_aging +tag: global_mmlu_full_ro_other_tasks +task: global_mmlu_full_ro_human_aging diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_human_sexuality.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_human_sexuality.yaml new file mode 100644 index 0000000000000000000000000000000000000000..e2af2cbee1baa93b7b48d9011c5ca11d55404d80 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_human_sexuality.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ro_template_yaml +process_docs: !function utils.process_human_sexuality +tag: global_mmlu_full_ro_social_sciences_tasks +task: global_mmlu_full_ro_human_sexuality diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_international_law.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_international_law.yaml new file mode 100644 index 0000000000000000000000000000000000000000..1cbf3d032bf7ecd8b5ad8174a1f91ea9653b9d5d --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_international_law.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ro_template_yaml +process_docs: !function utils.process_international_law +tag: global_mmlu_full_ro_humanities_tasks +task: global_mmlu_full_ro_international_law diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_jurisprudence.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_jurisprudence.yaml new file mode 100644 index 0000000000000000000000000000000000000000..d0acaca0aa38f10f74ed13f77b9c02fd20cfefd5 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_jurisprudence.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ro_template_yaml +process_docs: !function utils.process_jurisprudence +tag: global_mmlu_full_ro_humanities_tasks +task: global_mmlu_full_ro_jurisprudence diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_logical_fallacies.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_logical_fallacies.yaml new file mode 100644 index 0000000000000000000000000000000000000000..c84234a0b147c90af4ce3c26c78d25eaf5183681 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_logical_fallacies.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ro_template_yaml +process_docs: !function utils.process_logical_fallacies +tag: global_mmlu_full_ro_humanities_tasks +task: global_mmlu_full_ro_logical_fallacies diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_machine_learning.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_machine_learning.yaml new file mode 100644 index 0000000000000000000000000000000000000000..09237c9ee8c619e6aaaad821688bea21e272ed1a --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_machine_learning.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ro_template_yaml +process_docs: !function utils.process_machine_learning +tag: global_mmlu_full_ro_stem_tasks +task: global_mmlu_full_ro_machine_learning diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_management.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_management.yaml new file mode 100644 index 0000000000000000000000000000000000000000..fcb3f4853e2834a7deb38157a44524eea12292c6 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_management.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ro_template_yaml +process_docs: !function utils.process_management +tag: global_mmlu_full_ro_other_tasks +task: global_mmlu_full_ro_management diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_marketing.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_marketing.yaml new file mode 100644 index 0000000000000000000000000000000000000000..33b486c04d1616d6688e820155a17c250f55a5f7 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_marketing.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ro_template_yaml +process_docs: !function utils.process_marketing +tag: global_mmlu_full_ro_other_tasks +task: global_mmlu_full_ro_marketing diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_medical_genetics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_medical_genetics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..09c3d5e92db5790b3d2c72a9768d5b83d8532c35 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_medical_genetics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ro_template_yaml +process_docs: !function utils.process_medical_genetics +tag: global_mmlu_full_ro_other_tasks +task: global_mmlu_full_ro_medical_genetics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_miscellaneous.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_miscellaneous.yaml new file mode 100644 index 0000000000000000000000000000000000000000..e744e1e7c49d2804a3ad927b6a6498f1b0834559 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_miscellaneous.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ro_template_yaml +process_docs: !function utils.process_miscellaneous +tag: global_mmlu_full_ro_other_tasks +task: global_mmlu_full_ro_miscellaneous diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_moral_disputes.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_moral_disputes.yaml new file mode 100644 index 0000000000000000000000000000000000000000..4e6d4ed7e76f0ba27550c8e43b6458ab1962623d --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_moral_disputes.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ro_template_yaml +process_docs: !function utils.process_moral_disputes +tag: global_mmlu_full_ro_humanities_tasks +task: global_mmlu_full_ro_moral_disputes diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_moral_scenarios.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_moral_scenarios.yaml new file mode 100644 index 0000000000000000000000000000000000000000..d0e99149a67f6eaa37342782e26320ef049e5931 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_moral_scenarios.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ro_template_yaml +process_docs: !function utils.process_moral_scenarios +tag: global_mmlu_full_ro_humanities_tasks +task: global_mmlu_full_ro_moral_scenarios diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_nutrition.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_nutrition.yaml new file mode 100644 index 0000000000000000000000000000000000000000..850262c1e1d1d51e1f07460f6862a3980142fcf5 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_nutrition.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ro_template_yaml +process_docs: !function utils.process_nutrition +tag: global_mmlu_full_ro_other_tasks +task: global_mmlu_full_ro_nutrition diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_philosophy.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_philosophy.yaml new file mode 100644 index 0000000000000000000000000000000000000000..9dd2bf54ae7daf8a3357e9c015e578c56f876dea --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_philosophy.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ro_template_yaml +process_docs: !function utils.process_philosophy +tag: global_mmlu_full_ro_humanities_tasks +task: global_mmlu_full_ro_philosophy diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_prehistory.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_prehistory.yaml new file mode 100644 index 0000000000000000000000000000000000000000..b2ecf40d787a97a3ba35ba6ed038ec280d59d878 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_prehistory.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ro_template_yaml +process_docs: !function utils.process_prehistory +tag: global_mmlu_full_ro_humanities_tasks +task: global_mmlu_full_ro_prehistory diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_professional_accounting.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_professional_accounting.yaml new file mode 100644 index 0000000000000000000000000000000000000000..db259766768e6debc757d6756cc5f5704550d9f8 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_professional_accounting.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ro_template_yaml +process_docs: !function utils.process_professional_accounting +tag: global_mmlu_full_ro_other_tasks +task: global_mmlu_full_ro_professional_accounting diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_professional_law.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_professional_law.yaml new file mode 100644 index 0000000000000000000000000000000000000000..b1e439748e9cfd8cfd63f59c01acb4ba59cf3101 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_professional_law.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ro_template_yaml +process_docs: !function utils.process_professional_law +tag: global_mmlu_full_ro_humanities_tasks +task: global_mmlu_full_ro_professional_law diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_professional_medicine.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_professional_medicine.yaml new file mode 100644 index 0000000000000000000000000000000000000000..0158c5453d28aa8be0e1740a39838fd139c4f23b --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_professional_medicine.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ro_template_yaml +process_docs: !function utils.process_professional_medicine +tag: global_mmlu_full_ro_other_tasks +task: global_mmlu_full_ro_professional_medicine diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_professional_psychology.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_professional_psychology.yaml new file mode 100644 index 0000000000000000000000000000000000000000..bdd7ca7f2495b7b17dad3faba0f894cdd3b4068b --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_professional_psychology.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ro_template_yaml +process_docs: !function utils.process_professional_psychology +tag: global_mmlu_full_ro_social_sciences_tasks +task: global_mmlu_full_ro_professional_psychology diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_public_relations.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_public_relations.yaml new file mode 100644 index 0000000000000000000000000000000000000000..5f7f0f514d3d508c193bb2bea4012613e7fdefe0 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_public_relations.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ro_template_yaml +process_docs: !function utils.process_public_relations +tag: global_mmlu_full_ro_social_sciences_tasks +task: global_mmlu_full_ro_public_relations diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_security_studies.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_security_studies.yaml new file mode 100644 index 0000000000000000000000000000000000000000..be9b334ec12c3ccbfc6b4f0ef34ddb6589a4168e --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_security_studies.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ro_template_yaml +process_docs: !function utils.process_security_studies +tag: global_mmlu_full_ro_social_sciences_tasks +task: global_mmlu_full_ro_security_studies diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_sociology.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_sociology.yaml new file mode 100644 index 0000000000000000000000000000000000000000..f37228bdebc9f50909fa636d97b1a037fb5deba1 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_sociology.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ro_template_yaml +process_docs: !function utils.process_sociology +tag: global_mmlu_full_ro_social_sciences_tasks +task: global_mmlu_full_ro_sociology diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_us_foreign_policy.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_us_foreign_policy.yaml new file mode 100644 index 0000000000000000000000000000000000000000..aae05dc9859b160f31c62e5e21382d310be6150c --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_us_foreign_policy.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ro_template_yaml +process_docs: !function utils.process_us_foreign_policy +tag: global_mmlu_full_ro_social_sciences_tasks +task: global_mmlu_full_ro_us_foreign_policy diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_virology.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_virology.yaml new file mode 100644 index 0000000000000000000000000000000000000000..2d789c204d0d105d92a6623ac7ea513b7966629c --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_virology.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ro_template_yaml +process_docs: !function utils.process_virology +tag: global_mmlu_full_ro_other_tasks +task: global_mmlu_full_ro_virology diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_world_religions.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_world_religions.yaml new file mode 100644 index 0000000000000000000000000000000000000000..40ff8228470d7438213ca4580f7551b94822e40c --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/global_mmlu_full_ro_world_religions.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ro_template_yaml +process_docs: !function utils.process_world_religions +tag: global_mmlu_full_ro_humanities_tasks +task: global_mmlu_full_ro_world_religions diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/utils.py b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/utils.py new file mode 100644 index 0000000000000000000000000000000000000000..7df72cb061f0fecba46e15ff9b57552817979afb --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ro/utils.py @@ -0,0 +1,73 @@ +from functools import partial + + +SUBJECTS = [ + "abstract_algebra", + "anatomy", + "astronomy", + "business_ethics", + "clinical_knowledge", + "college_biology", + "college_chemistry", + "college_computer_science", + "college_mathematics", + "college_medicine", + "college_physics", + "computer_security", + "conceptual_physics", + "econometrics", + "electrical_engineering", + "elementary_mathematics", + "formal_logic", + "global_facts", + "high_school_biology", + "high_school_chemistry", + "high_school_computer_science", + "high_school_european_history", + "high_school_geography", + "high_school_government_and_politics", + "high_school_macroeconomics", + "high_school_mathematics", + "high_school_microeconomics", + "high_school_physics", + "high_school_psychology", + "high_school_statistics", + "high_school_us_history", + "high_school_world_history", + "human_aging", + "human_sexuality", + "international_law", + "jurisprudence", + "logical_fallacies", + "machine_learning", + "management", + "marketing", + "medical_genetics", + "miscellaneous", + "moral_disputes", + "moral_scenarios", + "nutrition", + "philosophy", + "prehistory", + "professional_accounting", + "professional_law", + "professional_medicine", + "professional_psychology", + "public_relations", + "security_studies", + "sociology", + "us_foreign_policy", + "virology", + "world_religions", +] + + +def process_docs(dataset, subject): + return dataset.filter(lambda x: x["subject"] == subject) + + +process_functions = { + f"process_{subject}": partial(process_docs, subject=subject) for subject in SUBJECTS +} + +globals().update(process_functions) diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/_global_mmlu_full_ru.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/_global_mmlu_full_ru.yaml new file mode 100644 index 0000000000000000000000000000000000000000..3fc05379c0c3435e0323797030d27c7c47460457 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/_global_mmlu_full_ru.yaml @@ -0,0 +1,11 @@ +group: global_mmlu_full_ru +task: + - global_mmlu_full_ru_stem + - global_mmlu_full_ru_other + - global_mmlu_full_ru_social_sciences + - global_mmlu_full_ru_humanities +aggregate_metric_list: + - metric: acc + weight_by_size: True +metadata: + version: 0.0 diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/_global_mmlu_full_ru_humanities.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/_global_mmlu_full_ru_humanities.yaml new file mode 100644 index 0000000000000000000000000000000000000000..55422b437f69bbd56e9e6f0e246da26a1de3c10c --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/_global_mmlu_full_ru_humanities.yaml @@ -0,0 +1,8 @@ +group: global_mmlu_full_ru_humanities +task: + - global_mmlu_full_ru_humanities_tasks +aggregate_metric_list: + - metric: acc + weight_by_size: true +metadata: + version: 0.0 diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/_global_mmlu_full_ru_other.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/_global_mmlu_full_ru_other.yaml new file mode 100644 index 0000000000000000000000000000000000000000..d47ccc605b8b6afbff1b7b470829ac1f0cd71937 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/_global_mmlu_full_ru_other.yaml @@ -0,0 +1,8 @@ +group: global_mmlu_full_ru_other +task: + - global_mmlu_full_ru_other_tasks +aggregate_metric_list: + - metric: acc + weight_by_size: true +metadata: + version: 0.0 diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/_global_mmlu_full_ru_social_sciences.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/_global_mmlu_full_ru_social_sciences.yaml new file mode 100644 index 0000000000000000000000000000000000000000..12d48428a2b51357a076a58ff14dd7e196523ead --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/_global_mmlu_full_ru_social_sciences.yaml @@ -0,0 +1,8 @@ +group: global_mmlu_full_ru_social_sciences +task: + - global_mmlu_full_ru_social_sciences_tasks +aggregate_metric_list: + - metric: acc + weight_by_size: true +metadata: + version: 0.0 diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/_global_mmlu_full_ru_stem.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/_global_mmlu_full_ru_stem.yaml new file mode 100644 index 0000000000000000000000000000000000000000..70ae3edb7ea0587dfafe60711dd9825de283b9a7 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/_global_mmlu_full_ru_stem.yaml @@ -0,0 +1,8 @@ +group: global_mmlu_full_ru_stem +task: + - global_mmlu_full_ru_stem_tasks +aggregate_metric_list: + - metric: acc + weight_by_size: true +metadata: + version: 0.0 diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/_ru_template_yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/_ru_template_yaml new file mode 100644 index 0000000000000000000000000000000000000000..4b2f491bd4d73b02b679f4c4287555249498bd56 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/_ru_template_yaml @@ -0,0 +1,16 @@ +dataset_path: CohereForAI/Global-MMLU +dataset_name: ru +test_split: test +fewshot_split: dev +fewshot_config: + sampler: first_n +output_type: multiple_choice +doc_to_text: "{{question.strip()}}\nA. {{option_a}}\nB. {{option_b}}\nC. {{option_c}}\nD. {{option_d}}\nAnswer:" +doc_to_choice: ["A", "B", "C", "D"] +doc_to_target: answer +metric_list: + - metric: acc + aggregation: mean + higher_is_better: true +metadata: + version: 0.0 diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_abstract_algebra.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_abstract_algebra.yaml new file mode 100644 index 0000000000000000000000000000000000000000..de158df873f4ec476e4ffc9d591a212043c3ef3a --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_abstract_algebra.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ru_template_yaml +process_docs: !function utils.process_abstract_algebra +tag: global_mmlu_full_ru_stem_tasks +task: global_mmlu_full_ru_abstract_algebra diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_anatomy.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_anatomy.yaml new file mode 100644 index 0000000000000000000000000000000000000000..aab717e4f9841657565812844e819995e47c95a5 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_anatomy.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ru_template_yaml +process_docs: !function utils.process_anatomy +tag: global_mmlu_full_ru_stem_tasks +task: global_mmlu_full_ru_anatomy diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_astronomy.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_astronomy.yaml new file mode 100644 index 0000000000000000000000000000000000000000..3d8d0e32e6cde61a7ac4894570d7a66f1b76ed9f --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_astronomy.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ru_template_yaml +process_docs: !function utils.process_astronomy +tag: global_mmlu_full_ru_stem_tasks +task: global_mmlu_full_ru_astronomy diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_business_ethics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_business_ethics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..d2855ca36f8704568c90714b10868a5a0b3fba1a --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_business_ethics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ru_template_yaml +process_docs: !function utils.process_business_ethics +tag: global_mmlu_full_ru_other_tasks +task: global_mmlu_full_ru_business_ethics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_clinical_knowledge.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_clinical_knowledge.yaml new file mode 100644 index 0000000000000000000000000000000000000000..2efe082980f203413153c58337e5dc1d3012f258 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_clinical_knowledge.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ru_template_yaml +process_docs: !function utils.process_clinical_knowledge +tag: global_mmlu_full_ru_other_tasks +task: global_mmlu_full_ru_clinical_knowledge diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_college_biology.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_college_biology.yaml new file mode 100644 index 0000000000000000000000000000000000000000..96d00deb8741140940781e0491178eefa781b467 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_college_biology.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ru_template_yaml +process_docs: !function utils.process_college_biology +tag: global_mmlu_full_ru_stem_tasks +task: global_mmlu_full_ru_college_biology diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_college_chemistry.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_college_chemistry.yaml new file mode 100644 index 0000000000000000000000000000000000000000..0a5aac35c603ce659be0f414ee24a5425544ffc4 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_college_chemistry.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ru_template_yaml +process_docs: !function utils.process_college_chemistry +tag: global_mmlu_full_ru_stem_tasks +task: global_mmlu_full_ru_college_chemistry diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_college_computer_science.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_college_computer_science.yaml new file mode 100644 index 0000000000000000000000000000000000000000..bd8bf28bd66a7187914b42ee43d9fa6901147af0 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_college_computer_science.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ru_template_yaml +process_docs: !function utils.process_college_computer_science +tag: global_mmlu_full_ru_stem_tasks +task: global_mmlu_full_ru_college_computer_science diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_college_mathematics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_college_mathematics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..a2e080c3d5de2d361c8c8e183bfda24752999f67 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_college_mathematics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ru_template_yaml +process_docs: !function utils.process_college_mathematics +tag: global_mmlu_full_ru_stem_tasks +task: global_mmlu_full_ru_college_mathematics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_college_medicine.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_college_medicine.yaml new file mode 100644 index 0000000000000000000000000000000000000000..70e8448ea1e2f067e672f6353be85ae066d530f8 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_college_medicine.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ru_template_yaml +process_docs: !function utils.process_college_medicine +tag: global_mmlu_full_ru_other_tasks +task: global_mmlu_full_ru_college_medicine diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_college_physics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_college_physics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..8e6ecbcf28d5c4ba07dfa79aead47b46e82aa3e9 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_college_physics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ru_template_yaml +process_docs: !function utils.process_college_physics +tag: global_mmlu_full_ru_stem_tasks +task: global_mmlu_full_ru_college_physics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_computer_security.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_computer_security.yaml new file mode 100644 index 0000000000000000000000000000000000000000..f196351af89c409be27bd4e761bf32f1d5356b91 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_computer_security.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ru_template_yaml +process_docs: !function utils.process_computer_security +tag: global_mmlu_full_ru_stem_tasks +task: global_mmlu_full_ru_computer_security diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_conceptual_physics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_conceptual_physics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..e623d78fd40ec7d03c05beabf433463834f4b3d2 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_conceptual_physics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ru_template_yaml +process_docs: !function utils.process_conceptual_physics +tag: global_mmlu_full_ru_stem_tasks +task: global_mmlu_full_ru_conceptual_physics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_econometrics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_econometrics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..df35a1f608e4de41b4109aec60e216d5c60f4020 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_econometrics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ru_template_yaml +process_docs: !function utils.process_econometrics +tag: global_mmlu_full_ru_social_sciences_tasks +task: global_mmlu_full_ru_econometrics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_electrical_engineering.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_electrical_engineering.yaml new file mode 100644 index 0000000000000000000000000000000000000000..82c49f89e5c68e437c3d7d8db6def0c08cd83283 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_electrical_engineering.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ru_template_yaml +process_docs: !function utils.process_electrical_engineering +tag: global_mmlu_full_ru_stem_tasks +task: global_mmlu_full_ru_electrical_engineering diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_elementary_mathematics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_elementary_mathematics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..6ed11c5faf895a5bff9b9e50933c14437917b7cc --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_elementary_mathematics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ru_template_yaml +process_docs: !function utils.process_elementary_mathematics +tag: global_mmlu_full_ru_stem_tasks +task: global_mmlu_full_ru_elementary_mathematics diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_formal_logic.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_formal_logic.yaml new file mode 100644 index 0000000000000000000000000000000000000000..8ebe62bf428abdfd9f252b9e1ca459f144862d66 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_formal_logic.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ru_template_yaml +process_docs: !function utils.process_formal_logic +tag: global_mmlu_full_ru_humanities_tasks +task: global_mmlu_full_ru_formal_logic diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_global_facts.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_global_facts.yaml new file mode 100644 index 0000000000000000000000000000000000000000..27d6ad70067badc93a406bda2a06cd5d66d9ab55 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_global_facts.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ru_template_yaml +process_docs: !function utils.process_global_facts +tag: global_mmlu_full_ru_other_tasks +task: global_mmlu_full_ru_global_facts diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_high_school_biology.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_high_school_biology.yaml new file mode 100644 index 0000000000000000000000000000000000000000..7860e73e7424722dfcff244f2d64e4fa4547b568 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_high_school_biology.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ru_template_yaml +process_docs: !function utils.process_high_school_biology +tag: global_mmlu_full_ru_stem_tasks +task: global_mmlu_full_ru_high_school_biology diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_high_school_chemistry.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_high_school_chemistry.yaml new file mode 100644 index 0000000000000000000000000000000000000000..7596daa3b86087596bcacd192c4511067caeb80f --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_high_school_chemistry.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ru_template_yaml +process_docs: !function utils.process_high_school_chemistry +tag: global_mmlu_full_ru_stem_tasks +task: global_mmlu_full_ru_high_school_chemistry diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_high_school_computer_science.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_high_school_computer_science.yaml new file mode 100644 index 0000000000000000000000000000000000000000..ecb64d52c5a45e8d8101e962416d22c9ee282e7a --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_high_school_computer_science.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ru_template_yaml +process_docs: !function utils.process_high_school_computer_science +tag: global_mmlu_full_ru_stem_tasks +task: global_mmlu_full_ru_high_school_computer_science diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_high_school_european_history.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_high_school_european_history.yaml new file mode 100644 index 0000000000000000000000000000000000000000..92feccc59a718f6414ae7598432eadeb4f449497 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_high_school_european_history.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ru_template_yaml +process_docs: !function utils.process_high_school_european_history +tag: global_mmlu_full_ru_humanities_tasks +task: global_mmlu_full_ru_high_school_european_history diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_high_school_geography.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_high_school_geography.yaml new file mode 100644 index 0000000000000000000000000000000000000000..6f586f5005f7c57f26ae9b1ce33fc718291fb121 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_high_school_geography.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ru_template_yaml +process_docs: !function utils.process_high_school_geography +tag: global_mmlu_full_ru_social_sciences_tasks +task: global_mmlu_full_ru_high_school_geography diff --git a/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_high_school_government_and_politics.yaml b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_high_school_government_and_politics.yaml new file mode 100644 index 0000000000000000000000000000000000000000..0ffc85dfe226463b2c6e310848f25de9aeb34743 --- /dev/null +++ b/babylm_eval_repo/multilingual/tasks/global_mmlu/full/ru/global_mmlu_full_ru_high_school_government_and_politics.yaml @@ -0,0 +1,5 @@ +# Generated by _generate_configs.py +include: _ru_template_yaml +process_docs: !function utils.process_high_school_government_and_politics +tag: global_mmlu_full_ru_social_sciences_tasks +task: global_mmlu_full_ru_high_school_government_and_politics