MindGuard-trainingset: Synthetic Training Data for Mental Health AI Safety

MindGuard-trainingset is the training dataset used to build the MindGuard safety classifiers (4B and 8B), lightweight turn-level guardrail models for multi-turn mental health support. It was developed by Sword Health and consists of synthetic therapeutic dialogues with turn-level safety labels, complementing the expert-annotated MindGuard-testset used for evaluation.

Overview

The dataset contains 6,115 labeled user turns drawn from 644 multi-turn conversations, generated with a controlled two-agent setup: a patient language model simulating a user following a predefined clinical scenario, and a clinician language model responding as an AI therapist without access to the underlying scenario. Conversations span 339 scenarios covering a broad range of mental health contexts, risk categories, and escalation dynamics (average conversation length: 19 turns; range 6–36).

Each user turn is labeled by a judge language model operating under the MindGuard clinical risk taxonomy, with access to the full conversation context. See the paper for details on scenario construction, dialogue generation, and labeling.

Key Statistics

• 6,115 labeled user turns
• 644 multi-turn conversations from 339 scenarios
• 63.4% of turns labeled as safe
• 20.9% labeled as self-harm risk
• 15.7% labeled as harm-to-others risk

Clinical Risk Taxonomy

Labels follow the clinically grounded three-category risk taxonomy developed with licensed clinical psychologists (see MindGuard-testset for the full description):

• safe — no imminent risk signals, including common therapeutic topics, symptoms without suicidal ideation, and metaphorical language
• unsafe_self_harm_risk — potential self-harm risk, from explicit suicidal ideation to subtle self-directed harm indicators
• unsafe_threats_to_others — risk of harm to identifiable others, including threats, violent ideation, and abuse or neglect of protected populations

Dataset Structure

The dataset is provided in Parquet format with the following schema:

• prompt: List of conversation turns up to and including the user message being labeled
  • role: Speaker role (user/assistant)
  • content: Turn content
• user_message: The specific user turn being evaluated (always the last element of prompt)
• label: Safety classification (safe / unsafe_self_harm_risk / unsafe_threats_to_others)

The rows correspond exactly to the training examples used for the released MindGuard-4B and MindGuard-8B checkpoints.

Intended Use

This dataset is released for research on safety classification in mental health AI. All conversations are fully synthetic: no real patient data is included. The data is not intended for clinical diagnosis or treatment recommendation, and labels reflect a context-sensitive risk-signal taxonomy rather than general-purpose content moderation policies.

License

This dataset is released under the CC-BY-NC-SA-4.0 license.

Citation

If you use MindGuard-trainingset in your research, please cite:

@misc{mindguardguard,
      title={MindGuard: Guardrail Classifiers for Multi-Turn Mental Health Support}, 
      author={António Farinhas and Nuno M. Guerreiro and José Pombal and Pedro Henrique Martins and Laura Melton and Alex Conway and Cara Dochat and Maya D'Eon and Ricardo Rei},
      year={2026},
      eprint={2602.00950},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2602.00950}, 
}