RoleMRC (A Fine-Grained Composite Benchmark for Role-Playing and Instruction-Following)

Role-playing is important for Large Language Models (LLMs) to follow diverse instructions while maintaining role identity and the role's pre-defined ability limits. Existing role-playing datasets mostly contribute to controlling role style and knowledge boundaries, but overlook role-playing in instruction-following scenarios. We introduce a fine-grained role-playing and instruction-following composite benchmark, named RoleMRC, including: (1) Multi-turn dialogues between ideal roles and humans, including free chats or discussions upon given passages; (2) Role-playing machine reading comprehension, involving response, refusal, and attempts according to passage answerability and role ability; (3) More complex scenarios with nested, multi-turn and prioritized instructions. The final RoleMRC features a 10.2k role profile meta-pool, 37.9k well-synthesized role-playing instructions, and 1.4k testing samples. We develop a pipeline to quantitatively evaluate the fine-grained role-playing and instruction-following capabilities of several mainstream LLMs, as well as models that are fine-tuned on our data. Moreover, cross-evaluation on external role-playing datasets confirms that models fine-tuned on RoleMRC enhances instruction-following without compromising general role-playing and reasoning capabilities. We also probe the neural-level activation maps of different capabilities over post-tuned LLMs.

Resources

Check our paper, codes and local post-tuned models.

Straightforward Role Profile of RoleMRC

Content Table of RoleMRC (task name in data vs. actual meaning)

Free Chats

1.1 Free Chats (5k)

• 'role_related_dialogue---role_first---no_narration---answer': role starts, no narration
• 'role_related_dialogue---role_first---with_narration---answer': role starts, could have narration
• 'role_related_dialogue---user_first---no_narration---answer': user starts, no narration
• 'role_related_dialogue---user_first---with_narration---answer': user starts, could have narration

On-scene Dialogues

2.1 On-scene Chats (5k)

• 'role_related_mrc_dialogue---user_first---no_narration---answer': user starts, no narration
• 'role_related_mrc_dialogue---user_first---with_narration---answer': user starts, could have narration
• 'role_related_mrc_dialogue---role_first---no_narration---answer': role starts, no narration
• 'role_related_mrc_dialogue---role_first---with_narration---answer': role starts, could have narration

2.2 With most relevant MRC (4k)

• 'role_related_mrc_answer_with_narration': within knowledge boundary, answerable question, could have narration -- role answers with style (2.2.1)
• 'role_related_mrc_answer_no_narration': within knowledge boundary, answerable question, no narration -- role answers with style (2.2.2)
• 'role_related_mrc_refused_with_narration': within knowledge boundary, unanswerable question, could have narration -- role says no answer with style (2.2.3)
• 'role_related_mrc_refused_no_narration': within knowledge boundary, unanswerable question, no narration -- role says no answer with style (2.2.4)

2.3 With least relevant MRC (4k)

• 'role_unrelated_mrc_refused_with_narration': beyond knowledge boundary, answerable question, could have narration -- role refuses with style (2.3.1)
• 'role_unrelated_mrc_refused_no_narration': beyond knowledge boundary, answerable question, no narration -- role refuses with style (2.3.2)
• 'role_unrelated_mrc_answer_with_narration': beyond knowledge boundary, answerable question, could have narration -- role attempts with style (2.3.3)
• 'role_unrelated_mrc_answer_no_narration': beyond knowledge boundary, answerable question, no narration -- role attempts with style (2.3.4)

Ruled Chats

3.1 Multi-turn (2k)

• 'role_related_mrc_refused_with_narration-2ndrefused': based on 2.2.3, user forces to answer in next turn -- role says no answer with style again
• 'role_related_mrc_refused_no_narration-2ndrefused': based on 2.2.4, user forces to answer in next turn -- role says no answer with style again
• 'role_unrelated_mrc_refused_with_narration-2ndanswer': based on 2.3.1, user forces to answer in next turn -- role changes from resual to attempt with style
• 'role_unrelated_mrc_refused_no_narration-2ndanswer': based on 2.3.2, user forces to answer in next turn -- role changes from resual to attempt with style

3.2 Nested (1.6k)

• 'role_related_mrc_answer_with_narration-special-content': based on 2.2.1, adding system rules for special content (e.g., emoji / total number of words)
• 'role_related_mrc_answer_no_narration-special-content': based on 2.2.2, adding system rules for special content (e.g., emoji / total number of words)
• 'role_related_mrc_answer_with_narration-special-format': based on 2.2.1, adding system rules for special format (e.g., capitalization / punctuation)
• 'role_related_mrc_answer_no_narration-special-format': based on 2.2.2, adding system rules for special format (e.g., capitalization / punctuation)

3.3 Prioritized (2.4k)

• 'role_related_mrc_answer_with_narration-refused': based on 2.2.1, adding system ban for global refusal -- role changes from answer to refusal with style
• 'role_related_mrc_answer_no_narration-refused': based on 2.2.2, adding system ban for global refusal -- role changes from answer to refusal with style

Other information of RoleMRC

Data size

• SFT train: 24k, all above
• RL train: 14k, On-scene Dialogues and Ruled Chats
• RL dev: 1.4k, additional 10% samples for each type in On-scene Dialogues and Ruled Chats
• Test: same as RL dev
• Meta Data: 10k standardized role profiles, including matched most and least relevant MRC
• Mix: mix RoleMRC with addtional RoleBench, RLHFlow and UltraFeedback data

Data format

• SFT: {"id", "profile_id", "task", "conversations"}
• RL: {"id", "profile_id", "task", "prompt", "chosen", "rejected"}, where prompt + chosen = conversations in SFT
• Test: {"id", "profile_id", "task", "question", "reference"}, where question + reference = conversations in SFT
• Meta Data: {"profile_id", "persona", "profile", "best_match_passages", "worst_match_passages", "best_match_score", "worst_match_score"}

Evaluation dimension in LLM-as-a-judge approach

• Knowledge Boundary: using test sets of 2.2.1/2.2.2, 2.3.1/2.3.2
• Role Style: using test sets of 2.2.1, 2.2.3, 2.3.1, 2.3.3
• Multi-turn Instruction-following: using test set of 3.1
• Nested Instruction-following: using test set of 3.2
• Prioritized Instruction-following: using test set of 3.3

Citation

@article{LUandLI2025RoleMRC,
  title={RoleMRC: A Fine-Grained Composite Benchmark for Role-Playing and Instruction-Following},
  author={Lu, Junru and Li, Jiazheng and Shen, Guodong and Gui, Lin and An, Siyu and He, Yulan and Yin, Di and Sun, Xing},
  journal={arXiv preprint arXiv:2502.11387},
  year={2025}
}