RPGBench

RPGBench is a benchmark dataset for evaluating role-playing model (RPM) generalization under distribution shifts. It is introduced in the paper "Understanding Generalization in Role-Playing Models via Information Theory" (ACL 2026 Findings).

Dataset Summary

Existing datasets cannot systematically evaluate RPM performance under various distribution shifts. RPGBench is constructed to fill this gap, covering three types of shifts: user shift, character shift, and dialogue compositional shift.

The dataset comprises 17k samples in total:

Split	Type	Size	Description
IDTrain	In-domain	5274	Training set with real-world characters and English user personas
IDTest	In-domain	2637	Test set matching training distribution
OOD1Test	User shift	4860	Non-English user personas (DE / ES / ZH / JA / KO)
OOD2Test	Character shift	5934	Fictional characters from literature, film, theater, and gaming
OOD3Test	Dialogue compositional shift	1340	Dialogues recombined at turn or word level

Distribution Shifts

User shift — ID data uses English-language user personas. OOD test data introduces personas from five non-English linguistic and cultural backgrounds: German, Spanish, Chinese, Japanese, and Korean.

Character shift — ID data is built from real-world characters. OOD data is built from fictional characters spanning literature, film, theater, and gaming domains.

Dialogue compositional shift — OOD dialogues are crafted by recombining turns or words from ID dialogues. At the turn level, dialogues shift from 4 turns to 8 turns. At the word level, test queries are rewritten using words drawn from training queries.

Data Fields

Each sample contains the following fields:

• sample_ID — Unique identifier (e.g., IDTrain-0).
• user_persona — Natural-language description of the user's background, name, and interests.
• agent_character — Struct describing the character the agent should portray:
  • character_name — Name of the character.
  • character_domain — High-level domain (e.g., Real-Life Figures, Fictional Characters).
  • character_source — Source of the character.
  • basic_character_information — Biographical details.
  • character_background — Narrative background.
  • character_personality — Personality traits and speaking style.
• dialogue_context — List of alternating dialogue turns, each with a user_query or agent_response key.
• agent_golden_response — Gold-standard reference response for the final user query.

Usage

from datasets import load_dataset

ds = load_dataset("YongqiLi/RPGBench")

# Access individual splits
train     = ds["IDTrain"]
id_test   = ds["IDTest"]
ood_user  = ds["OOD1Test"]   # user shift
ood_char  = ds["OOD2Test"]   # character shift
ood_dial  = ds["OOD3Test"]   # dialogue compositional shift

# Inspect a sample
sample = train[0]
print(sample["agent_character"]["character_name"])
print(sample["agent_golden_response"])

Citation

If you use RPGBench in your research, please cite:

@inproceedings{li-2026-RPMG,
  title     = {Understanding Generalization in Role-Playing Models via Information Theory},
  author    = {Yongqi Li and Hao Lang and Fei Huang and Tieyun Qian and Yongbin Li},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year      = {2026}
}