somosnlp-hackathon-2026/chaskiayuda-dataset-dpo
收藏资源简介:
ChasquiAyuda DPO 是一个用于训练西班牙语人道主义响应助手的数据集,专门针对秘鲁背景。该数据集旨在解决秘鲁发生灾害时,受影响人群面临的双重危机:灾害本身以及关于他们有权获得何种援助以及如何获取援助的虚假信息。通用人工智能助手通常会给出模糊或不适用于秘鲁背景的答案。ChasquiAyuda 训练一个模型,以提供针对秘鲁的可操作、共情且文化适宜的响应,包含真实的电话号码(如INDECI 115)、真实的援助计划(如Programa Contigo, SIS)以及可验证的人道主义标准(如Sphere, CHS 2024)。数据集遵循标准DPO格式,包含提示(prompt)、优选回复(chosen)和拒绝回复(rejected)。它结合了8个不同来源的数据构建而成,包括国际标准、秘鲁官方来源、真实危机背景来源以及针对2026年厄尔尼诺现象的数据。数据集总共包含约620个偏好对,涵盖了多种危机类型,并按照90%训练集和10%测试集进行划分。
ChasquiAyuda DPO is a preference dataset (DPO) for training a humanitarian response assistant in Spanish, specialized in the Peruvian context. The dataset aims to address the dual crisis faced by affected people in Peru during disasters: the disaster itself and misinformation about what aid they are entitled to and how to access it. Generic AI assistants often give vague or inapplicable answers for the Peruvian context. ChasquiAyuda trains a model to respond with actionable, empathetic, and culturally appropriate information for Peru, including real phone numbers (e.g., INDECI 115), real programs (e.g., Programa Contigo, SIS), and verifiable humanitarian standards (e.g., Sphere, CHS 2024). The dataset follows the standard DPO format, containing a prompt, a chosen response, and a rejected response. It was constructed by combining 8 different sources, including international standards, Peruvian official sources, real crisis context sources, and data for the 2026 El Niño phenomenon. The dataset contains approximately 620 preference pairs, covering various types of crises, and is split into 90% training set and 10% test set.




