MistraMystic: Conversational Personality Model 🌊✨

Welcome to MistraMystic—a conversational model fine-tuned from Mistral-7B v0.3, capturing nuanced personality traits that make AI interactions feel more authentic and relatable. Whether it’s about balancing conscientious responses or tapping into empathetic reflections, MistraMystic is here to explore the depths of the human-like personality spectrum.

Model Name: MistraMystic

Architecture: Mistral-7B v0.3
Training Objective: Personality-Enhanced Conversational AI
Training Dataset: Fine-tuned on conversational data to reflect Big 5 personality traits.
- JIC: Journal Intensive Conversations dataset
Training Duration: 4-5 days on A100 GPU (training parameters can be found in appendix of the paper)

Why "MistraMystic"?

The name "MistraMystic" combines the mystique of deep conversation with Mistral's adaptability. Designed to capture the essence of personality through the Big 5 OCEAN traits, MistraMystic works to reflect the nuances of human interactions within its AI responses. The result? A model that speaks with more than just words—it reflects aspects of personality, adding richness and realism to every interaction.

Scope of Applications

MistraMystic is crafted for a range of applications where understanding personality-driven conversation is essential. Here’s what it’s especially good for:

Conversational Agents: Engage users with relatable and personality-driven conversations.
Text Generation: Generate human-like text for articles, chats, and creative writing with a personal touch.
Question-Answering: Answer questions with a flair of personality, making responses more relatable.
Educational and Therapy Bots: Assist in applications where personality-sensitive responses can improve user engagement and retention.

Intended Use

MistraMystic is built for those aiming to inject personality into conversational systems, whether it’s for customer service bots, therapy support, or just plain fun AI companions. It’s particularly suited to applications where capturing nuances like openness, agreeableness, and neuroticism (yes, even those angsty replies!) can enhance user experience.

Data and Training

The model has been trained on an extensive conversational dataset. Our goal was to align model responses with intrinsic personality traits, enabling MistraMystic to tailor its tone and style depending on conversational context. More information on the dataset will be shared soon.

Results

Personality Evaluation on EleutherAI/lm-evaluation-harness (OCEAN Personality Benchmark)

Model	Description	Openness	Conscientiousness	Extraversion	Agreeableness	Neuroticism	Average
Mistral 7B v0.3	Zero-shot	0.8360	0.6390	0.5140	0.8160	0.5350	0.6680
MistraMystic	Fine-tuned on Conversational Data	0.9340	0.8260	0.6250	0.9530	0.5700	0.7816

MistraMystic demonstrates notable improvements across all Big 5 traits.

Performance and Limitations

While MistraMystic brings vibrant and personality-driven conversations to the table, it does have limitations:

Personality Representation: MistraMystic is trained for personality alignment, so it may sacrifice some general knowledge capabilities in favor of personality-specific responses. A detailed evaluation will be updated soon.
Sensitive Topics: Despite strong filtering, caution is advised when deploying in high-stakes environments.
Computational Load: The Mistral 7B backbone requires substantial resources, which may limit deployment in real-time settings without sufficient hardware.

Ethical Considerations

We made sure to avoid toxic or inappropriate dialogues by tagging any dialogue with over 25% toxic utterances for separate review. Ethical considerations are a priority, and MistraMystic was designed with responsible AI practices in mind. For details on ethical data practices, see the Appendix (coming soon!).

Future Updates

Stay tuned for more information on MistraMystic!

Citation

@inproceedings{pal-etal-2025-beyond,
    title = "Beyond Discrete Personas: Personality Modeling Through Journal Intensive Conversations",
    author = "Pal, Sayantan  and
      Das, Souvik  and
      Srihari, Rohini K.",
    editor = "Rambow, Owen  and
      Wanner, Leo  and
      Apidianaki, Marianna  and
      Al-Khalifa, Hend  and
      Eugenio, Barbara Di  and
      Schockaert, Steven",
    booktitle = "Proceedings of the 31st International Conference on Computational Linguistics",
    month = jan,
    year = "2025",
    address = "Abu Dhabi, UAE",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2025.coling-main.470/",
    pages = "7055--7074",
    abstract = "Large Language Models (LLMs) have significantly improved personalized conversational capabilities. However, existing datasets like Persona Chat, Synthetic Persona Chat, and Blended Skill Talk rely on static, predefined personas. This approach often results in dialogues that fail to capture human personalities' fluid and evolving nature. To overcome these limitations, we introduce a novel dataset with around 400,000 dialogues and a framework for generating personalized conversations using long-form journal entries from Reddit. Our approach clusters journal entries for each author and filters them by selecting the most representative cluster, ensuring that the retained entries best reflect the author`s personality. We further refine the data by capturing the Big Five personality traits{---}openness, conscientiousness, extraversion, agreeableness, and neuroticism{---}ensuring that dialogues authentically reflect an individual`s personality. Using Llama 3 70B, we generate high-quality, personality-rich dialogues grounded in these journal entries. Fine-tuning models on this dataset leads to an 11{\%} improvement in capturing personality traits on average, outperforming existing approaches in generating more coherent and personality-driven dialogues."
}

choco58
/

MistraMystic