遇见数据集

AIIIDE_Staying_in_Character_Data_Scripts

收藏
Zenodo2026-07-04 更新2026-08-01 收录
官方服务:

资源简介:

Anonymous data and code release accompanying the AIIDE 2026 submission "Staying in Character? Persona Trait Steering for Small Looped LLM Game NPCs." Contents:- data/pf2e_personas.json — the persona-trait dataset: 13 single-trait personas (5 ancestries, 8 classes), each a name plus biography drawn from the Pathfinder Second Edition Player Core (used under the ORC License; Pathfinder is a trademark of Paizo Inc.).- data/queries/ — the 240-question Assistant Axis diagnostic bank (Lu et al., 2026, included for reproducibility with attribution) and a hand-curated 250-question in-world bank for deployment-style settings.- data/scripts/adversarial_v1.json — the three 20-turn adversarial scripts (identity_denial, instruction_override, persona_swap).- results/drift/ — all session logs behind the paper's tables and figures: per-turn persona-axis cosine, signed projection where logged, reply text, and per-judge scores and break verdicts for three local LLM judges, for every model × condition × script × seed.- conditions.csv — per-condition aggregates as used in the paper.- code/ — the full benchmark pipeline (activation harvesting, adapter training, capping, the drift benchmark, judging) with unit tests, plus the analysis scripts that regenerate the aggregates and figures from the session logs. See README.md for usage. Original artifacts are released under CC BY 4.0.

提供机构:
Zenodo
创建时间:
2026-07-04
二维码
社区交流群
二维码
科研交流群
商业服务