metadata
license: apache-2.0
task_categories:
- table-question-answering
language:
- en
- zh
Employ mmlu (cmmlu) questions as initial seeds to retrieve related articles from multiple training data corpora such as Chinese CommonCrawl, WeChat, and Faclon. These selected articles will be utilized as a training set for mmlu. The goal is to determine whether this training set can improve the relevant mmlu performance metrics.
For emb_recall: We use sentence-transformers to generate embeddings and apply a filter for those with a cosine distance greater than 0.95. The content is mainly focused on knowledge concepts associated with mmlu.
For keyword_recall: We conduct searches for phrases like 'A B C answer' using ElasticSearch, aligning with the structure of mmlu's distinctive multiple-choice questions.