init

LM-Evaluation-Harness-240310/lm_eval/tasks/belebele/belebele_tgl_Latn.yaml

+"fewshot_split": "tgl_Latn"
+"include": "_default_template_yaml"
+"task": "belebele_tgl_Latn"
+"test_split": "tgl_Latn"

LM-Evaluation-Harness-240310/lm_eval/tasks/belebele/belebele_tha_Thai.yaml

+"fewshot_split": "tha_Thai"
+"include": "_default_template_yaml"
+"task": "belebele_tha_Thai"
+"test_split": "tha_Thai"

LM-Evaluation-Harness-240310/lm_eval/tasks/belebele/belebele_tir_Ethi.yaml

+"fewshot_split": "tir_Ethi"
+"include": "_default_template_yaml"
+"task": "belebele_tir_Ethi"
+"test_split": "tir_Ethi"

LM-Evaluation-Harness-240310/lm_eval/tasks/belebele/belebele_tsn_Latn.yaml

+"fewshot_split": "tsn_Latn"
+"include": "_default_template_yaml"
+"task": "belebele_tsn_Latn"
+"test_split": "tsn_Latn"

LM-Evaluation-Harness-240310/lm_eval/tasks/belebele/belebele_tso_Latn.yaml

+"fewshot_split": "tso_Latn"
+"include": "_default_template_yaml"
+"task": "belebele_tso_Latn"
+"test_split": "tso_Latn"

LM-Evaluation-Harness-240310/lm_eval/tasks/belebele/belebele_tur_Latn.yaml

+"fewshot_split": "tur_Latn"
+"include": "_default_template_yaml"
+"task": "belebele_tur_Latn"
+"test_split": "tur_Latn"

LM-Evaluation-Harness-240310/lm_eval/tasks/belebele/belebele_ukr_Cyrl.yaml

+"fewshot_split": "ukr_Cyrl"
+"include": "_default_template_yaml"
+"task": "belebele_ukr_Cyrl"
+"test_split": "ukr_Cyrl"

LM-Evaluation-Harness-240310/lm_eval/tasks/belebele/belebele_urd_Arab.yaml

+"fewshot_split": "urd_Arab"
+"include": "_default_template_yaml"
+"task": "belebele_urd_Arab"
+"test_split": "urd_Arab"

LM-Evaluation-Harness-240310/lm_eval/tasks/belebele/belebele_urd_Latn.yaml

+"fewshot_split": "urd_Latn"
+"include": "_default_template_yaml"
+"task": "belebele_urd_Latn"
+"test_split": "urd_Latn"

LM-Evaluation-Harness-240310/lm_eval/tasks/belebele/belebele_uzn_Latn.yaml

+"fewshot_split": "uzn_Latn"
+"include": "_default_template_yaml"
+"task": "belebele_uzn_Latn"
+"test_split": "uzn_Latn"

LM-Evaluation-Harness-240310/lm_eval/tasks/belebele/belebele_vie_Latn.yaml

+"fewshot_split": "vie_Latn"
+"include": "_default_template_yaml"
+"task": "belebele_vie_Latn"
+"test_split": "vie_Latn"

LM-Evaluation-Harness-240310/lm_eval/tasks/belebele/belebele_war_Latn.yaml

+"fewshot_split": "war_Latn"
+"include": "_default_template_yaml"
+"task": "belebele_war_Latn"
+"test_split": "war_Latn"

LM-Evaluation-Harness-240310/lm_eval/tasks/belebele/belebele_wol_Latn.yaml

+"fewshot_split": "wol_Latn"
+"include": "_default_template_yaml"
+"task": "belebele_wol_Latn"
+"test_split": "wol_Latn"

LM-Evaluation-Harness-240310/lm_eval/tasks/belebele/belebele_xho_Latn.yaml

+"fewshot_split": "xho_Latn"
+"include": "_default_template_yaml"
+"task": "belebele_xho_Latn"
+"test_split": "xho_Latn"

LM-Evaluation-Harness-240310/lm_eval/tasks/belebele/belebele_yor_Latn.yaml

+"fewshot_split": "yor_Latn"
+"include": "_default_template_yaml"
+"task": "belebele_yor_Latn"
+"test_split": "yor_Latn"

LM-Evaluation-Harness-240310/lm_eval/tasks/belebele/belebele_zho_Hans.yaml

+"fewshot_split": "zho_Hans"
+"include": "_default_template_yaml"
+"task": "belebele_zho_Hans"
+"test_split": "zho_Hans"

LM-Evaluation-Harness-240310/lm_eval/tasks/belebele/belebele_zho_Hant.yaml

+"fewshot_split": "zho_Hant"
+"include": "_default_template_yaml"
+"task": "belebele_zho_Hant"
+"test_split": "zho_Hant"

LM-Evaluation-Harness-240310/lm_eval/tasks/belebele/belebele_zsm_Latn.yaml

+"fewshot_split": "zsm_Latn"
+"include": "_default_template_yaml"
+"task": "belebele_zsm_Latn"
+"test_split": "zsm_Latn"

LM-Evaluation-Harness-240310/lm_eval/tasks/belebele/belebele_zul_Latn.yaml

+"fewshot_split": "zul_Latn"
+"include": "_default_template_yaml"
+"task": "belebele_zul_Latn"
+"test_split": "zul_Latn"

LM-Evaluation-Harness-240310/lm_eval/tasks/benchmarks/flan/_held_in_template_yaml

+output_type: generate_until
+test_split: null
+doc_to_choice: null
+metric_list:
+  - metric: exact_match
+    aggregation: mean
+    higher_is_better: true
+generation_kwargs:
+  until:
+    - "</s>"
+  do_sample: false
+  temperature: 0.0
+metadata:
+  version: 1.0