File Structure

• rolebench-eng/instruction-generalization, rolebench-eng/role-generalization, and rolebench-zh: All contain two subfolders: general and role_specific. Each subfolder has training data, testing data, and the RoleGPT baseline results for comparison.
• rolegpt baseline: A strong baseline for performance comparison. RoleGPT is Role Prompting using GPT for speaking style imitation. In the case of general instructions, we generate 6 responses using RoleGPT for each role-instruction pair; 5 of these serve as ground truth candidates, while the remaining one functions as a RoleGPT strong baseline.