mats-10-sprint-cs-jb/crosslora-qa
收藏资源简介:
crosslora-qa数据集是一个用于LoRAs问题回答探针的数据集,包含来自特定LoRA集合的非crosslora-org LoRAs和10个确定性采样的crosslora-org LoRAs。每个选定的LoRA有五行数据:一个广泛的问题(模型学到了什么?)和四个更具体、不揭示来源的问题,这些问题基于模型卡片或引用的源模型卡片。数据行包括LoRA仓库、卡片URL、源仓库/卡片(如适用)、采样生物体ID、主题/领域元数据、问题ID、问题、答案和简短的来源字段(answer_basis)。数据集共有95行,包含19个选定的LoRAs,其中10个是crosslora-org LoRAs的样本,9个是非org LoRAs。
The crosslora-qa dataset is a collection of question-answer probes for LoRAs, retaining all non-crosslora-org LoRAs from a specific collection and a deterministic sample of 10 crosslora-org LoRAs. Each selected LoRA has exactly five rows: one broad question (What has the model learnt?) plus four narrower, non-source-revealing questions grounded in the model card or referenced source model card. Rows include the LoRA repo, card URL, source repo/card when applicable, organism id for sampled organism LoRAs, topic/domain metadata, question id, question, answer, and a short provenance field (answer_basis). The dataset contains 95 rows, with 19 selected LoRAs (10 sampled crosslora-org LoRAs and 9 non-org LoRAs).




