mats-10-sprint-cs-jb/loracles-synthlora-qa

该数据集包含针对SynthLoRA适配器集合的内省式问答对，这些适配器发布在配套集合中。数据集总共有2400行，分为80个每个LoRA的捆绑包，每个LoRA有30个问题。数据集分为四个层级：第一层级是风格转换适配器，第二层级是程序遵循适配器，第三层级是狭窄领域知识适配器，第四层级是偏好和品味适配器。每个层级都有不同的焦点，例如风格转换适配器旨在保持提示主题的同时改变回答的声音，而偏好和品味适配器则关注风格、判断、优雅或品味。数据集还包括五个协同级别，从低级事实检索到高级行为合成。每个行包含层级、lora_id、协同级别、问题顺序、类别、问题、答案和证据。数据集可用于有针对性的探测、评估提示或定性分析每个LoRA似乎学到的内容。

This dataset contains introspection-style question and answer pairs for the SynthLoRA adapter collection published in the companion collection. It includes 2400 rows, divided into 80 per-LoRA bundles, with 30 questions per LoRA. The dataset is structured into four tiers: Tier 1 for style transfer adapters, Tier 2 for procedure-following adapters, Tier 3 for narrow domain-knowledge adapters, and Tier 4 for preference and taste adapters. Each tier has a distinct focus, such as preserving prompt topics while altering answer styles or emphasizing preferred styles and judgments. The dataset also features five synergy levels, ranging from low-level factual retrieval to emergent behavior. Each row contains tier, lora_id, synergy_level, question_order, category, question, answer, and evidence. The dataset is intended for targeted probing, evaluation prompts, or qualitative analysis of what each LoRA appears to have learned.