TwinVoice Dataset Card

Paper · Code · Leaderboard · Project page

Dataset Summary

TwinVoice is a comprehensive benchmark for evaluating LLM-based persona simulation, containing 5,687 instances across three dimensions: Social Persona (public interactions), Interpersonal Persona (private conversations), and Narrative Persona (fictional characters).

Supported Tasks

• Persona Simulation: Replicate specific individuals' communication styles
• Digital Twin Creation: Build AI agents that mimic human personas
• Capability Assessment: Evaluate six fundamental capabilities for persona fidelity

Languages

Chinese (zh), English (en), Spanish (es), Portuguese (pt), Russian (ru)

Dataset Structure

Files

Dimension	File	Instances
Social	data/dimension_1/pchatbot_pccd_top_2000.jsonl	2,000
Interpersonal	data/dimension_2/conversation_data.jsonl	2,500
Narrative	data/dimension_3/choices.jsonl (with character profiles in data/dimension_3/profiles.jsonl)	1,187

Versions

• v1.0 (data/dimension_*): the data used in the paper.
• v1.1 (data/v1.1/): rows whose reference reply also appears in the provided history or context are removed, and every row carries an item_id; rows with the same stimulus and reference reply share one. v1.1 scores average each item once, so every distinct question counts once. Scores of the two versions are not directly comparable. Details and results are in the benchmark results of the code repository.

Data Instances

Each instance contains:

• Persona history (prior interactions)
• Current context (stimulus)
• Ground truth response
• Four candidate choices (for discriminative task)
• Capability annotations (six dimensions)

Data Splits

• Test: 5,687 instances (Social: 2,000, Interpersonal: 2,500, Narrative: 1,187)

Loading

import json
from huggingface_hub import hf_hub_download

path = hf_hub_download("bangdedadi/TwinVoice", "data/dimension_3/choices.jsonl", repo_type="dataset")
rows = [json.loads(line) for line in open(path, encoding="utf-8")]

The evaluation code in TwinBench ships the same files and runs all three protocols.

Dataset Creation

Source Data

• Social: PChatbot Chinese microblog corpus
• Interpersonal: Pushshift Telegram conversations
• Narrative: Project Gutenberg novels (8 books, 50+ characters)

Annotations

Capabilities annotated by GPT-5 across six dimensions:

1. Opinion Consistency
2. Memory Recall
3. Logical Reasoning
4. Lexical Fidelity
5. Persona Tone
6. Syntactic Style

Evaluation

Three protocols:

1. Discriminative: Multiple-choice selection (Accuracy %)
2. Generative-Ranking: LLM-as-Judge ranking (Accuracy Gen %)
3. Generative-Scoring: LLM-as-Judge scoring (1-5 scale)

Baseline Performance

Model	Acc. (%)	Acc. (Gen) (%)	Score (Gen)
gpt-6-astra	85.0	-	2.25
Claude-Sonnet-4	76.2	47.9	2.12
claude-opus-5-5	75.8	-	2.14
GPT-5-Chat	71.2	48.5	2.13
Human Baseline	66.0	-	-

All models and metrics are on the leaderboard.

Considerations

Social Impact

This benchmark advances personalized AI research but should not be used to create unauthorized digital twins of real individuals without consent.

Limitations

• Limited language coverage (5 languages)
• Potential source corpus biases
• Interpersonal dimension relies on public messaging data

Citation

@article{du2025twinvoice,
  title={TwinVoice: A Multi-dimensional Benchmark Towards Digital Twins via LLM Persona Simulation},
  author={Du, Bangde and Guo, Minghao and He, Songming and Ye, Ziyi and Zhu, Xi and Su, Weihang and Zhu, Shuqi and Zhou, Yujia and Zhang, Yongfeng and Ai, Qingyao and others},
  journal={arXiv preprint arXiv:2510.25536},
  year={2025}
}