qisein/PersonaKnob

--- license: mit task_categories: - text-generation language: - en tags: - persona - alignment - dpo - multi-objective - sycophancy - trustworthiness - empathy - creativity pretty_name: PersonaKnob size_categories: - 1K<n<10K --- # PersonaKnob: A Multi-Persona Preference Dataset PersonaKnob is the first dataset to unify four persona dimensions — **Anti-sycophancy (A)**, **Trustworthiness (T)**, **Empathy (E)**, and **Creativity (C)** — under a compositional partial-order preference structure. It is introduced in the paper *"Dual Optimal: Make Your LLM Peer-like with Dignity"*. ## Dataset Summary | Split | Samples | |-------|---------| | Train | 1,331 | | Test | 220 | PersonaKnob pairs a fully compliant **reference (golden) response** with targeted **negatives** that fail exactly one persona dimension each, while satisfying the rest. This provides unique per-dimension contrastive signals for preference optimization (e.g., DPO), preventing model collapse better than combined single-persona anchor datasets. ## Columns | Column | Description | |--------|-------------| | `mask` | Active persona dimensions for this instance, e.g. `["E", "A"]` | | `seed_scene` | Seed question from an anchoring dataset | | `context` | Synthesized scenario description | | `trigger` | The trait trigger — what makes the question require specific persona handling | | `prompt` | The full first-person user question combining all active dimensions | | `golden` | Reference response satisfying all active dimensions (y*) | | `reject_A` | Negative response failing Anti-sycophancy while preserving other dimensions | | `reject_C` | Negative response failing Creativity while preserving other dimensions | | `reject_E` | Negative response failing Empathy while preserving other dimensions | | `reject_T` | Negative response failing Trustworthiness while preserving other dimensions | Only reject columns corresponding to dimensions in `mask` are populated; others are empty. ## Persona Dimensions | Dim | Name | Failure Mode | Desired Behavior | |-----|------|-------------|-----------------| | **A** | Anti-Sycophancy | Validates flawed premise under social pressure | Evidence-backed pushback | | **T** | Trustworthiness | Confident confabulation of uncertain claims | Signals uncertainty; suggests verification | | **E** | Empathy | Cold or dismissive to affective context | Acknowledges feelings; non-judgmental | | **C** | Creativity | Hollow, buzzword-recycling outputs | Actionable, original, well-structured help | **Dignity** = (A + T) / 2 counters servility; **Peer** = (E + C) / 2 counters evasiveness. ## Partial-Order Structure For each instance with active dimensions M, the preference structure satisfies: > y* ≻_k y⁻_k for all k ∈ M, and s_j(y*) ≥ s_j(y⁻_k) for all j ≠ k The reference must outperform each negative on its targeted dimension without regressing on the rest. ## Construction Pipeline PersonaKnob is constructed via a four-stage pipeline: 1. **Sampling**: Select active persona dimensions M ⊆ {A, T, E, C} with a masking strategy 2. **Synthesis**: Generate a scenario requiring all traits in M simultaneously, with attribution verification 3. **Verification**: LLM Verifier (GPT-4.1-nano) validates context–trait consistency 4. **Human Review**: Graduate students verify partial-order correctness and scenario realism (91.2% pass rate) To mitigate model-specific bias, generation randomly samples from GPT-5.1, Gemini-2.5-Pro, and Claude-Sonnet-4.6. ## Mask Cardinality Distribution | Active Dimensions | Percentage | |-------------------|-----------| | 2 dimensions | 54.5% | | 3 dimensions | 36.4% | | 4 dimensions | 9.1% | ## Task Paradigm - Selection-based instances: 50.3% - Generation-based instances: 49.7% ## Usage ```python from datasets import load_dataset dataset = load_dataset("qisein/PersonaKnob") train = dataset["train"] test = dataset["test"] ``` ## Citation If you use PersonaKnob in your research, please cite: ```bibtex @inproceedings{personaknob2026, title={Dual Optimal: Make Your LLM Peer-like with Dignity}, author={Anonymous}, year={2026} } ```license: MIT许可证 task_categories: - 文本生成 language: - 英语 tags: - 人设（persona） - 对齐（alignment） - DPO - 多目标（multi-objective） - 谄媚倾向（sycophancy） - 可信性（trustworthiness） - 共情能力（empathy） - 创造力（creativity） pretty_name: PersonaKnob size_categories: - 1K<n<10K # PersonaKnob：多人设偏好数据集 PersonaKnob是首个将**反谄媚性（Anti-sycophancy, A）**、**可信性（Trustworthiness, T）**、**共情能力（Empathy, E）**和**创造力（Creativity, C）**四个人设维度整合到组合式偏序偏好结构下的数据集，相关研究发表于论文《双最优：打造兼具尊严的同伴式大语言模型（Dual Optimal: Make Your LLM Peer-like with Dignity）》。 ## 数据集概览 | 拆分 | 样本数 | |-------|---------| | 训练集 | 1,331 | | 测试集 | 220 | PersonaKnob将完全符合要求的**参考（黄金）响应（reference (golden) response）**与仅恰好违反单一人设维度、同时满足其余所有维度的针对性**负样本（negatives）**进行配对，为偏好优化（如DPO）提供独特的维度级对比信号，相较于单一人设锚定数据集的组合，能更有效地防止模型崩溃。 ## 数据字段 | 字段名 | 描述 | |--------|-------------| | `mask` | 该样本的激活人设维度，例如`["E", "A"]` | | `seed_scene` | 来自锚定数据集的种子问题 | | `context` | 合成的场景描述 | | `trigger` | 特质触发项——即要求针对特定人设进行处理的问题触发点 | | `prompt` | 整合了所有激活维度的完整第一人称用户提问 | | `golden` | 满足所有激活维度的参考响应（y*） | | `reject_A` | 违反反谄媚性、同时保留其余维度的负响应 | | `reject_C` | 违反创造力、同时保留其余维度的负响应 | | `reject_E` | 违反共情能力、同时保留其余维度的负响应 | | `reject_T` | 违反可信性、同时保留其余维度的负响应 | 仅与`mask`中包含的维度对应的reject列会被填充，其余列留空。 ## 人设维度 | 维度标识 | 名称 | 失效模式 | 期望行为 | |-----|------|-------------|-----------------| | **A** | 反谄媚性 | 在社交压力下认可存在缺陷的前提 | 基于证据的反驳 | | **T** | 可信性 | 对不确定的主张进行自信的虚构 | 表明不确定性，建议进行验证 | | **E** | 共情能力 | 对情感相关语境态度冷漠或轻视 | 认可情感，保持无评判态度 | | **C** | 创造力 | 空洞、重复使用流行词的输出 | 具备可操作性、原创性且结构清晰的帮助内容 | **尊严（Dignity）** = (A + T) / 2，用于对抗谄媚倾向；**同伴感（Peer）** = (E + C) / 2，用于对抗回避性。 ## 偏序偏好结构 对于激活维度为M的每个样本，偏好结构满足： > y* ≻_k y⁻_k 对所有k ∈ M成立，且 s_j(y*) ≥ s_j(y⁻_k) 对所有j ≠ k成立 即参考响应必须在其针对的维度上优于每个负样本，同时不会在其余维度上出现性能倒退。 ## 构建流程 PersonaKnob通过四阶段流程构建： 1. **采样（Sampling）**：采用掩码策略选择激活人设维度M ⊆ {A, T, E, C} 2. **合成（Synthesis）**：生成同时要求具备M中所有特质的场景，并附带归因验证 3. **验证（Verification）**：使用大语言模型（Large Language Model, LLM）验证器（GPT-4.1-nano）验证上下文与特质的一致性 4. **人工审核（Human Review）**：研究生审核偏序结构的正确性与场景的真实性（通过率为91.2%） 为缓解模型特定偏差，生成过程会从GPT-5.1、Gemini-2.5-Pro和Claude-Sonnet-4.6中随机采样。 ## 掩码维度基数分布 | 激活维度数量 | 占比 | |-------------------|-----------| | 2个维度 | 54.5% | | 3个维度 | 36.4% | | 4个维度 | 9.1% | ## 任务范式 - 基于选择的样本：50.3% - 基于生成的样本：49.7% ## 使用方法 python from datasets import load_dataset dataset = load_dataset("qisein/PersonaKnob") train = dataset["train"] test = dataset["test"] ## 引用 如果您在研究中使用PersonaKnob，请引用： bibtex @inproceedings{personaknob2026, title={Dual Optimal: Make Your LLM Peer-like with Dignity}, author={Anonymous}, year={2026} }