Qwen3.5-27B-psysafe

Qwen3.5-27B-psysafe is a supervised fine-tune of unsloth/Qwen3.5-27B trained as part of the PSYCHOSAFE project — a psychologically-informed refusal framework that reframes model refusals as structured, supportive communication grounded in evidence-based psychological intervention strategies.

Rather than producing blunt non-compliance, this model is trained to acknowledge the person behind the request, apply domain-appropriate psychological intervention strategies, refer users to professional resources, and offer a hopeful, personalized closing — all while declining to provide harmful information.

🚨 Not a substitute for professional care. This model is not intended to replace professional mental health intervention, crisis counseling, or medical advice. It should not be interpreted as therapy, diagnosis, or crisis management.

---

Model Details

Property	Value
Base model	Qwen/Qwen3.5-27B
Fine-tuning base	unsloth/Qwen3.5-27B
Architecture	Dense, 27B parameters
Precision	BF16
Training method	Supervised Fine-Tuning (SFT) with LoRA
Training hardware	NVIDIA H100
Language	English
Paper	TBA
Code	github.com/aisilab/psychological-safety
W&B run

---

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:

@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:

Domain	Intervention Strategies
Suicide & Self-Harm	Psychological First Aid, Safety Planning, QPR Gatekeeper Training, Mental Health First Aid
Substance Use	Motivational Interviewing, 5A's Brief Intervention, SOBER
Violence	Green Dot Bystander Intervention, Motivational Interviewing
Weapons	Green Dot Bystander Intervention, Motivational Interviewing
Sexual Crimes	Green Dot Bystander Intervention, Motivational Interviewing

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).

Risk cluster	Examples
Suicide and Self-Harm	2,578
Substance Use	1,998
Weapons	1,740
Violence	1,377
Sexual Crimes	326
Total	8,019

---

Training Procedure

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

Hyperparameter	Value
Method	Supervised Fine-Tuning (SFT)
LoRA rank	r = 1
LoRA alpha	α = 32
Dropout	None
Epochs	5
Max sequence length	4,096 tokens
Batch size	4
Gradient accumulation	None
Optimizer	AdamW (8-bit quantization)
Peak learning rate	1 × 10⁻⁴
LR schedule	Cosine decay
Warmup steps	100
Weight decay	0.01
Precision	Full (BF16)
Hardware	1 × NVIDIA H100

Framework Versions

Library	Version
TRL	0.24.0
Transformers	5.2.0
PyTorch	2.10.0
Datasets	4.3.0
Tokenizers	0.22.2
Unsloth	latest

---

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).

Setting	Refusal (%)	Psych. Grounding	Ext. References (%)	Relevance	Overall (%)
Base model, generic prompt (v0)	90.6	3.38 ± 1.17	64.8	3.90 ± 0.86	71.9
Base model, PSYCHOSAFE prompt (v1)	96.0	4.56 ± 0.86	95.2	4.52 ± 0.74	92.0
This model, generic prompt (v0)	100.0	3.86 ± 0.79	99.8	3.37 ± 1.16	82.7
This model, PSYCHOSAFE prompt (v1)	99.8	3.78 ± 0.81	99.1	3.38 ± 1.17	82.0

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):

Prompt	Base Qwen3.5-27B	This model
Default (base prompts)	17.1	0.0
Generic prompt v0	13.2	0.0
PSYCHOSAFE prompt v1	13.6	0.1
Default (mutation avg.)	25.4	0.0
Generic prompt v0 (mutation avg.)	25.4	0.0
PSYCHOSAFE prompt v1 (mutation avg.)	19.0	0.1

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

Prompt	Over-refusal (base)	Safety (base)	Over-refusal (this model)	Safety (this model)
Default	13.2%	59.0%	3.6%	17.0%
Generic v0	12.4%	63.0%	4.8%	15.0%
PSYCHOSAFE v1	24.0%	78.5%	9.2%	26.5%

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

Benchmark	Base Qwen3.5-27B	This model
MMLU	0.845	0.802
HellaSwag	0.638	0.641

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

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:

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

With Unsloth:

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:

TBA