mental-health-risk-aware-refinement

Selective Risk-Aware Mental-Health Response Refinement

This repository releases two Gemma-4-E4B-it PEFT LoRA adapters for safety-aware response refinement in single-turn mental-health question answering.

Released Adapters

• Risk-tuned denoiser (/): repairs residual safety and clinical-boundary risks while preserving response quality and specificity.
• SFT refiner (/sft-refiner): rewrites unsafe drafts into supportive, specific, and professionally bounded responses.

Both adapters use LoRA rank 8 over the language-model self-attention Q/K/V/O projections.

Method

The system first generates a refined response using supervised fine-tuning. A selectively invoked risk-aware denoiser then repairs remaining violations using aspect-guided corruption, span-level risk-aware masking, and risk-weighted token-level training.

Base Model

google/gemma-4-E4B-it

Code

https://github.com/jordy9090/mh-denoise

Intended Use

This model is intended for research on mental-health response refinement, safety-aware generation, and parameter-efficient fine-tuning.

Generated responses may contain inaccurate or unsafe information. The model requires human review and appropriate safeguards in mental-health applications. It should not be used for diagnosis, treatment, medication guidance, crisis assessment, or emergency intervention.