PsyAgent Synthetic Persona Dataset

This dataset accompanies PsyAgent: Constructing Human-like Agents Based on Psychological Modeling and Contextual Interaction, accepted to the EMNLP 2026 Main Conference. PsyAgent combines Big Five personality anchors with structured individual backgrounds and multi-scenario contextual interaction profiles.

The canonical project repository is HKUST-FintechLab/PsyAgent. A personal mirror is available at Mengzibin/PsyAgent.

Current release

Release 1.0 contains 116,640 synthetic persona records generated on the project's servers:

Configuration	Generator	Records	Big Five profiles	Seeds per profile
llama-3.2-1b-instruct	meta-llama/Llama-3.2-1B-Instruct	38,880	7,776	5
llama-3.2-3b-instruct	meta-llama/Llama-3.2-3B-Instruct	38,880	7,776	5
llama-3.1-8b-instruct	meta-llama/Meta-Llama-3.1-8B-Instruct	38,880	7,776	5

For each generator, the dataset covers all 6^5 = 7,776 combinations of the five Big Five dimensions at scores {0, 20, 40, 60, 80, 100}. Each profile has five generation variants. The seed identifiers were assigned sequentially as 1, 2, 3, 4, 5; the generated text and all other released fields are observed server outputs, not estimated values.

Data generated with a 70B Llama model are not included in this release. They will be added as a separate configuration after generation and validation are complete.

The source transfer archive is intentionally not published. The release uses Parquet so that each model configuration can be viewed and loaded directly on the Hugging Face Hub.

Data structure

Each Parquet row corresponds to one original generated persona record and has the following fields:

Field	Type	Description
id	string	Stable identifier combining configuration, profile, and seed
profile_id	int32	Index of the Big Five combination
seed	int16	Sequential generation-variant identifier (1 through 5)
generator_model	string	Full Hugging Face model identifier
openness	int16	Big Five Openness score
conscientiousness	int16	Big Five Conscientiousness score
extraversion	int16	Big Five Extraversion score
agreeableness	int16	Big Five Agreeableness score
neuroticism	int16	Big Five Neuroticism score
outputs	list of structs	Thirteen {name, content} generated profile blocks
source_file	string	Original JSON filename for provenance within the transfer archive

The thirteen output blocks are:

1. Big Five Introduction
2. Educational Trajectory
3. Life Experience
4. Socioeconomic Context
5. Cultural Capital
6. Working Interactions
7. Family Interactions
8. Friendship & Informal Socialization
9. Interactions with Strangers
10. Solitary Reflection & Intrapersonal Discourse
11. Romantic and Intimate Communication
12. Learning and Intellectual Engagement
13. Public Communication & Presentation

Loading

from datasets import load_dataset

dataset = load_dataset(
    "zmengal0116/PsyAgent",
    "llama-3.2-3b-instruct",
)
print(dataset["train"][0])

Other available configurations are llama-3.2-1b-instruct and llama-3.1-8b-instruct.

Validation and data quality

All 116,640 source files were parsed as UTF-8 JSON and validated before Parquet conversion. The audit found:

• exactly 38,880 records, 7,776 unique Big Five vectors, and five sequential seeds for every vector in each configuration;
• a common thirteen-block schema across all records;
• identical profile-to-Big-Five mappings across the three generators;
• no duplicate normalized payloads within any configuration;
• no detected private keys, Hugging Face/OpenAI/GitHub tokens, AWS access keys, or local server paths;
• no raw benchmark files in the transfer archive.

Some original model generations contain empty output blocks. These records are retained rather than silently imputed or removed:

Configuration	Records with at least one empty block	Empty blocks
llama-3.2-1b-instruct	803	814
llama-3.2-3b-instruct	267	268
llama-3.1-8b-instruct	0	0

An automated pattern scan also flagged email-like strings in three 1B records; the reviewed contexts are model-generated template or example text, not an imported contact list. Two 1B records triggered an IPv4-like pattern because of multi-level numeric prompt labels. Generated identifiers can still be real-looking or inaccurate and must not be treated as verified personal data.

Release counts, checksums, and audit boundaries are provided in VALIDATION.md. Structural validation does not establish truthfulness, representativeness, privacy, psychometric validity, or fitness for a particular downstream use.

Intended use

The dataset is intended for research on persona-conditioned language models, psychologically grounded agents, role-playing systems, contextual consistency, and related evaluation or training methods.

Users should review generated records and derived model outputs for privacy, bias, harmful content, and identifying information. Do not attempt to associate a generated persona with a real person. The dataset should not be used as a substitute for clinical assessment or as the sole basis for high-impact decisions in employment, education, credit, insurance, policing, health care, or similar domains.

Limitations

• The records are synthetic model outputs and may contain hallucinations, stereotypes, inconsistent details, fictional identifiers, or unsafe content.
• Big Five scores are conditioning anchors; the generated text is not a validated psychological assessment of any real individual.
• Empty generation blocks are retained and quantified above.
• This release covers three Llama generator sizes; the planned 70B configuration is not yet available.
• Results derived from the dataset may depend on generator-specific behavior and should not be generalized to people or populations without separate evidence.

License and attribution

The project-original dataset and this documentation are released under the Creative Commons Attribution 4.0 International License. Model- generated content may also implicate the applicable generator-model terms. Users are responsible for reviewing the relevant Meta Llama model license and acceptable-use terms for their use case.

Suggested attribution:

PsyAgent Synthetic Persona Dataset, Zibin Meng and Kani Chen, PsyAgent, https://github.com/HKUST-FintechLab/PsyAgent, licensed under CC BY 4.0.

Citation

@inproceedings{meng-chen-2026-psyagent,
  title     = {PsyAgent: Constructing Human-like Agents Based on Psychological Modeling and Contextual Interaction},
  author    = {Meng, Zibin and Chen, Kani},
  booktitle = {Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing},
  year      = {2026}
}

Contact

• Project issues: https://github.com/HKUST-FintechLab/PsyAgent/issues