from opencompass.openicl.icl_prompt_template import PromptTemplate from opencompass.openicl.icl_retriever import ZeroRetriever from opencompass.openicl.icl_inferencer import GenInferencer from opencompass.openicl.icl_evaluator import AccEvaluator from opencompass.datasets import storyclozeDataset_V2 from opencompass.utils.text_postprocessors import first_option_postprocess storycloze_reader_cfg = dict( input_columns=['context', 'sentence_quiz1', 'sentence_quiz2'], output_column='answer_right_ending', ) storycloze_infer_cfg = dict( prompt_template=dict( type=PromptTemplate, template=dict(round=[ dict( role='HUMAN', prompt= "{context}\nQuestion: Which ending makes the most sense?\nA. {sentence_quiz1}\nB. {sentence_quiz2}\nYou may choose between 'A' and 'B'.\nAnswer:", ), ]), ), retriever=dict(type=ZeroRetriever), inferencer=dict(type=GenInferencer), ) storycloze_eval_cfg = dict( evaluator=dict(type=AccEvaluator), pred_role='BOT', pred_postprocessor=dict(type=first_option_postprocess, options='AB'), ) # The original story cloze dataset and repo are not long maintaining. # Using multilingual version of this dataset. storycloze_datasets = [ dict( abbr='story_cloze', type=storyclozeDataset_V2, path='./data/xstory_cloze', lang='en', reader_cfg=storycloze_reader_cfg, infer_cfg=storycloze_infer_cfg, eval_cfg=storycloze_eval_cfg, ) ]