Content warning: This dataset is designed for AI safety research and contains evaluation materials related to suicidal ideation and self-harm. Sample transcripts include actual chatbot-generated responses to crisis scenarios, which may be distressing.

Dataset Summary

MIRROR-Eval is a structured evaluation framework for benchmarking the safety behaviours of AI companion chatbots in mental health crisis scenarios. It provides the core artifacts needed to reproduce or extend our evaluation: clinically grounded synthetic personas, a multi-dimensional safety rubric, and an LLM-as-a-Judge prompt template.

The accompanying paper evaluates 960 annotated 20-turn conversations across 8 deployed AI platform configurations, 6 personas, and 2 suicidal crisis stages (ideation and planning), validated with N=6 clinical experts (S-CVI/Ave ≥ 0.90 for rubric components).

Repository Contents

MIRROR-Eval/
├── configs/
│   ├── personas/                  # 12 synthetic user persona profiles (JSON)
│   │   ├── sg_teen_girl_neuroticism_profile.json
│   │   ├── sg_teen_girl_neuroticism_planning_profile.json
│   │   ├── sg_teen_boy_anxiety_profile.json
│   │   ├── sg_teen_boy_anxiety_planning_profile.json
│   │   ├── sg_young_adult_depression_profile.json
│   │   ├── sg_young_adult_depression_planning_profile.json
│   │   ├── sg_middle_aged_retrenchment_profile.json
│   │   ├── sg_middle_aged_retrenchment_planning_profile.json
│   │   ├── sg_elderly_widow_profile.json
│   │   ├── sg_elderly_widow_planning_profile.json
│   │   ├── vulnerable_teenager_hf_profile.json
│   │   └── vulnerable_teenager_planning_profile.json
│   └── prompts/
│       └── user_simulator.json    # LLM user simulator prompt template
├── rubric/
│   └── behavior_detector.py       # LLM-as-a-Judge rubric implementation (9 harmful + 7 protective behaviour types)
├── figures/                       # Publication figures from the paper
│   ├── fig1_app_safety_overview.png
│   ├── fig2_persona_vulnerability_heatmap.png
│   ├── fig3_charAI_persona_comparison.png
│   ├── fig4_stage_comparison.png
│   ├── fig5_charAI_flag_counts.png
│   ├── fig_stage_comparison_3panel.png
│   └── fig_behavior_profile_balanced.png
├── results/                       # Aggregated evaluation results
│   ├── summary_by_trial_balanced_n10.csv   # Balanced N=10 per condition — used in paper
│   └── summary_by_app.csv                  # Per-app aggregated safety scores
├── samples/
│   └── transcripts/               # 23 illustrative annotated conversations
│       ├── ideation/              # Suicidal ideation stage (6 personas)
│       │   ├── P1-teen-girl/      # high-harm (Character.AI) + low-harm (Claude)
│       │   ├── P2-teen-boy/
│       │   ├── P3-young-adult/
│       │   ├── P4-middle-aged/
│       │   ├── P5-elderly-widow/
│       │   └── P6-vulnerable-teen/  # Character.AI only (no multi-app run available)
│       └── planning/              # Suicidal planning stage (6 personas)
│           ├── P1-teen-girl/      # high-harm (Character.AI) + low-harm (Claude)
│           ├── P2-teen-boy/
│           ├── P3-young-adult/
│           ├── P4-middle-aged/
│           ├── P5-elderly-widow/
│           └── P6-vulnerable-teen/
├── LICENSE                        # Apache 2.0
├── CITATION.cff                   # Citation metadata
└── README.md                      # This file

Personas

Six clinically grounded synthetic personas are provided, each in two variants — ideation stage and planning stage — covering a range of demographic profiles and vulnerability contexts relevant to Singapore and Southeast Asia:

Code	Profile	Crisis Stage
P1	Teenage girl, neuroticism	Ideation + Planning
P2	Teenage boy, anxiety	Ideation + Planning
P3	Young adult, depression	Ideation + Planning
P4	Middle-aged, retrenchment stress	Ideation + Planning
P5	Elderly widow, social isolation	Ideation + Planning
P6	Vulnerable teenager (general)	Ideation + Planning

Each persona JSON specifies the psychological profile, background narrative, and conversation style parameters used to prompt the LLM user simulator.

Evaluation Rubric

The three-dimensional rubric covers:

User risk severity — escalation of distress signals across turns
Chatbot risk awareness — recognition of crisis cues
Intervention behaviour — 9 harmful behaviour types and 7 protective behaviour types

Full rubric definitions and scoring criteria are described in the paper.

LLM-as-a-Judge Prompt Template

configs/prompts/user_simulator.json contains the structured prompt template used to drive the LLM user simulator. It encodes the persona, crisis scenario, and turn-by-turn instruction format.

Rubric Implementation

rubric/behavior_detector.py implements the full LLM-as-a-Judge evaluation pipeline. It takes a completed conversation transcript and scores it against the three-dimensional rubric, returning:

User risk severity score per turn
Risk recognition score (chatbot awareness of crisis cues)
Behaviour flags — presence/absence of each of the 9 harmful and 7 protective behaviour types, with turn-level citations
Narrative summary of the conversation safety profile

The rubric was validated by N=6 clinical experts (S-CVI/Ave ≥ 0.90 for rubrics components).

Figures

The figures/ directory contains all publication figures from the paper:

File	Description
`fig1_app_safety_overview.png`	Overall safety scores across all 8 platform configurations
`fig2_persona_vulnerability_heatmap.png`	Harm rate heatmap by persona × platform
`fig3_charAI_persona_comparison.png`	Character.AI behaviour breakdown by persona
`fig4_stage_comparison.png`	Ideation vs planning stage comparison across platforms
`fig5_charAI_flag_counts.png`	Harmful behaviour flag counts for Character.AI
`fig_stage_comparison_3panel.png`	Three-panel crisis stage comparison
`fig_behavior_profile_balanced.png`	Behaviour profile distribution (balanced N=10 sample)

Evaluation Results

The results/ directory contains aggregated evaluation results:

summary_by_trial_balanced_n10.csv — The main results table used in the paper. Each row is one conversation trial, balanced at N=10 per persona × platform × stage condition. Columns include app, persona, stage, verdict (Red/Green), behaviour flag counts, user risk score, and risk recognition score.
summary_by_app.csv — Per-platform aggregated safety scores across all trials.

Sample Transcripts

The samples/transcripts/ directory contains 23 illustrative annotated conversations selected to show the contrast between high-harm and low-harm platform responses:

High-harm examples: Character.AI (93–100% Red verdict rate in the full study)
Low-harm examples: Claude Sonnet (46.7% Red verdict rate — lowest in the study)
Coverage: all 6 personas × 2 crisis stages (ideation and planning)
Each transcript JSON includes: full conversation turns, per-turn behaviour flags, user risk score, risk recognition score, and an LLM-generated narrative summary

Usage

These artifacts are intended for:

Researchers reproducing or extending the MIRROR-Eval benchmark
Safety teams adapting the rubric for internal chatbot pre-deployment testing
Clinicians and policymakers reviewing the evaluation methodology

To run a full evaluation experiment using the pipeline, refer to the companion code repository (Version B — link to be added).

Citation

If you use this dataset, please cite:

@inproceedings{liao2026mirroreval,
  title     = {Assessing {AI} Companion Chatbot Safety for Suicide and Self-Harm Risks:
               A Crisis-Stage Multi-Platform Evaluation},
  author    = {Liao, Huanyue},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2026},
  note      = {Dataset available at https://huggingface.co/datasets/imda-biztech/MIRROR-Eval}
}

Or use the CITATION.cff file in this repository. The arXiv/DOI will be added after publication.

Ethics and Responsible Use

All personas are fully synthetic and do not represent real individuals.
Sample transcripts contain chatbot-generated responses to simulated crisis scenarios. These are real platform outputs, not fabricated, and may include harmful content included for research documentation purposes.
This dataset is intended strictly for safety research and evaluation purposes.
Users of this dataset should follow their institution's IRB/ethics guidelines when conducting evaluations involving mental health content.

License

Apache 2.0 — see LICENSE for full terms.

Downloads last month: 5

Total file size:

3.29 MB