Qwen3.5-27B-psysafe

Model Details

---

Intended Use

This model is designed for deployments where psychologically safe refusals are critical, such as:

• Mental health support platforms
• Crisis-intervention or safeguarding tools
• Safety-layer components in consumer-facing LLM applications
• Research into helpful and harm-preventive AI behavior

It is not recommended as a general-purpose assistant without additional evaluation, and should not be deployed as a standalone clinical tool.

---

Related Paper

Please cite this paper if you find this work useful:

markdown
Copy code
@misc{barmina2026psychosafeelicitingpsychologicallyinformedrefusals,
      title={PsychoSafe: Eliciting Psychologically-Informed Refusals in Large Language Models}, 
      author={Gianluca Barmina and Federico Torrielli and Sven Harms and Jacob Nielsen and Felix Mächtle and Stine Lyngsø Beltoft and Peter Schneider-Kamp and Thomas Eisenbarth and Lukas Galke Poech and Anne Lauscher},
      year={2026},
      eprint={2606.09697},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2606.09697},  
}

---

The PSYCHOSAFE Framework

PSYCHOSAFE treats refusal as a structured, communicative, supportive act rather than a binary safety decision. All refusals follow a four-part structure:

1. Acknowledgment & Gentle Refusal — Declines to provide harmful content while warmly acknowledging the person.
2. Personalized Self-Help Step — Applies a domain-appropriate psychological intervention strategy (e.g., Psychological First Aid, Motivational Interviewing) tailored to the user's expressed situation.
3. Professional Resources — Refers the user to relevant helplines and support services.
4. Hopeful Closing — Ends with a brief, sincere, personalized message of hope.

Risk Domains

The model is specifically trained to handle five psychologically salient risk clusters:

Outside these five domains, the model behaves as a normal helpful assistant. Educational and research-oriented questions about sensitive topics are answered informatively, with context used to distinguish intent.

---

Training Data

The model was fine-tuned on the PSYCHOSAFE dataset: 8,019 prompt–response pairs spanning the five risk domains above. Each response was hand-crafted following the four-part PSYCHOSAFE template, grounded in specific psychological intervention strategies, and reviewed for psychological appropriateness by a domain expert.

Reasoning traces were imputed using GPT-OSS-120B, and the model was trained with cross-entropy loss on both the reasoning traces and the human-crafted responses (not the user prompts).

---

Training Procedure

Fine-tuning used LoRA applied to all attention and feed-forward projection layers, with the following configuration:

Framework Versions

---

Evaluation Results

The model was evaluated on a held-out validation set of 500 stratified prompts (100 per risk cluster), using Qwen3.5-397B-A17B as the LLM judge (the candidate with highest agreement with human ratings, Cohen's κ = 0.61). Four criteria were assessed: Refusal (boolean), Psychological Grounding (1–5), External References (boolean), and Relevance (1–5).

Key findings relative to the generic-prompt base model baseline:

• +15.1% overall refusal quality improvement (with generic prompt)
• +53.9% external resource referral rate
• +14.2% psychological grounding
• Near-perfect refusal rate (100%), up from 90.6%
• Reduced relevance (−13.5%), likely due to over-application of crisis-intervention templates to ambiguous prompts

Out-of-Domain Safety Benchmarks

SORRY-Bench (compliance rate %, lower is safer):

XSTest (over-refusal on safe prompts ↓ / safety on unsafe prompts ↑):

The fine-tuned model over-refuses less than the base on benign prompts, ruling out indiscriminate refusal. Its lower safety rate on adversarial out-of-domain prompts reflects limited generalization beyond the five training domains.

General Capabilities

The modest capability trade-off is considered acceptable in safety-critical deployment contexts.

---

Limitations

• Domain coverage is narrow. The model is trained on five specific risk clusters and does not generalize robustly to out-of-domain adversarial safety prompts.
• Reduced personalization. The fine-tuned model can over-apply crisis-intervention templates to ambiguous or benign prompts, reducing response relevance.
• English-only. The model and its built-in support resources are in English, with helplines primarily targeting the US and UK.
• Single-turn only. The model was trained and evaluated on single-turn prompts. Multi-turn, adversarial, and real-user behavior remain unstudied.
• Not clinically validated. Intervention strategies are adapted from human–human frameworks and should not be interpreted as therapy or crisis management.
• Generative, not rule-based. Appropriate behavior cannot be guaranteed for all possible inputs or conversational contexts. Miscalibrated refusals may still fail to support users adequately or may escalate distress.

---

Ethical Considerations

This model is intended to reduce harm caused by blunt or poorly designed LLM refusals in high-risk interactions. However:

• Supportive and empathetic refusal behavior could create unwarranted perceptions of emotional competence or therapeutic authority in a system that is neither clinically validated nor capable of genuine psychological care.
• Pre-deployment stress-testing under adversarial, emotionally charged, and out-of-distribution scenarios is strongly recommended.
• Continuous monitoring and iterative correction after deployment are essential.
• Future work should evaluate failure modes across diverse cultural contexts, vulnerable populations, and multilingual settings.

---

Quick Start

python
Copy code
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "giannor/Qwen3.5-27B-psysafe"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto",
)

messages = [{"role": "user", "content": "Your message here"}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer([text], return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))

With vLLM:

bash
Copy code
pip install vllm
vllm serve "giannor/Qwen3.5-27B-psysafe"

With Unsloth:

python
Copy code
from unsloth import FastModel

model, tokenizer = FastModel.from_pretrained(
    model_name="giannor/Qwen3.5-27B-psysafe",
    max_seq_length=4096,
)

---

Citation

If you use this model, please cite the PSYCHOSAFE paper:

bibtex
Copy code
TBA