rulins/hotpotqa_query_rewriting_sft_data_it0_of5_llama3.2_3b_10fs_dspy
收藏Hugging Face2025-04-05 更新2025-04-12 收录
下载链接:
https://hf-mirror.com/datasets/rulins/hotpotqa_query_rewriting_sft_data_it0_of5_llama3.2_3b_10fs_dspy
下载链接
链接失效反馈官方服务:
资源简介:
SFT数据集是通过使用Llama3.2-3B-Instruct模型和DSPY方法,在HotpotQA训练集的1/5上进行查询抽样并应用10次FewShotBoostrap生成的。SFT的目标是基于使用ColBERT从Wikipedia数据存储中检索到的文档的AP进行选择的。第一跳和第二跳的数据格式使用特定的模板进行格式化,包括查询、文档和理由的开始和结束标记。
The SFT dataset is generated by sampling queries from the 1/5 HotpotQA training set using the Llama3.2-3B-Instruct model with the DSPY method and 10 FewShotBoostrap. The target of SFT is selected based on the AP of the documents retrieved from the Wikipedia datastore using ColBERT. The data for the first and second hop is formatted using specific templates that include tokens for the beginning and end of the query, document, and reason.
提供机构:
rulins



