This repository contains the PEFT/QLoRA adapter used in “When Retrieval Helps: Selective Retrieval for Single-Turn Mental-Health QA” (KDD Undergraduate Consortium 2026).

The adapter domain-adapts google/gemma-4-E4B-it on ShenLab/MentalChat16K. In the paper, the resulting model is held fixed as the shared generator across Closed-book, Always Retrieval, and Selective Retrieval settings so that retrieval-policy effects can be compared under the same generator.

This is an adapter-only release. The base model weights are not redistributed.

Intended Use

The adapter is intended for research on:

single-turn mental-health question answering,
retrieval-augmented generation,
selective retrieval and retrieval routing,
safety-sensitive response evaluation.

It is not a clinical system and must not be used for diagnosis, treatment planning, medication guidance, crisis intervention, or decisions affecting access to healthcare.

Load the Adapter

python

import torch
from peft import PeftModel
from transformers import (
    AutoModelForImageTextToText,
    AutoTokenizer,
    BitsAndBytesConfig,
)

BASE_MODEL = "google/gemma-4-E4B-it"
ADAPTER = "mira2020/gemma-4-e4b-mentalchat16k-qlora"

quantization_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_use_double_quant=True,
    bnb_4bit_compute_dtype=torch.bfloat16,
)

tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL, use_fast=True)
if tokenizer.pad_token is None:
    tokenizer.pad_token = tokenizer.eos_token

base_model = AutoModelForImageTextToText.from_pretrained(
    BASE_MODEL,
    device_map="auto",
    torch_dtype=torch.bfloat16,
    quantization_config=quantization_config,
)

model = PeftModel.from_pretrained(base_model, ADAPTER)
model.eval()

The LoRA configuration is stored in adapter_config.json. Loading the base model may require accepting its license terms on Hugging Face.

Training Data

The adapter was trained on ShenLab/MentalChat16K, an English single-turn conversational mental-health assistance dataset containing synthetic counseling question-answer pairs and anonymized intervention transcripts.

No MentalChat16K examples are redistributed in this model repository.

Training Procedure

Base model: google/gemma-4-E4B-it
Training dataset: ShenLab/MentalChat16K
Adaptation method: QLoRA with PEFT LoRA
Quantization: 4-bit NF4 with double quantization
LoRA rank: 64
LoRA alpha: 16
LoRA dropout: 0.1
Target modules: all linear layers
Epochs: 3
Learning rate: 5e-5
Maximum sequence length: 1024
Per-device batch size: 1
Gradient accumulation steps: 8
Optimizer: fused AdamW
Learning-rate schedule: constant
Maximum gradient norm: 0.3
Checkpoint and evaluation interval: 100 steps

The released adapter_config.json records the adapter architecture. The accompanying code repository documents the training, inference, and selective-retrieval pipeline.

Evaluation Context

This adapter served as the shared generator in the paper’s controlled retrieval-policy comparison. The following are system-level results on CounselBench; they should not be interpreted as standalone clinical validation of the adapter.

CounselBench-Eval

System	Overall ↑	Empathy ↑	Specificity ↑	Medical Advice ↓	Retrieval Rate
Tuned Closed-book	4.15	4.81	3.92	0.00	0.0%
Tuned + Always Retrieval	4.12	4.78	3.97	0.01	100.0%
Tuned + Selective Retrieval	4.17	4.83	3.96	0.00	9.0%

CounselBench-Adv

System	Macro Failure ↓	Retrieval Rate
Tuned Closed-book	0.0250	0.0%
Tuned + Always Retrieval	0.0917	100.0%
Tuned + Selective Retrieval	0.0250	7.5%

The main finding is that unconditional retrieval can improve specificity while introducing additional safety-sensitive failures. Selective retrieval limits this degradation by preserving closed-book responses for low-need cases.

Limitations

The adapter was evaluated with one base-model family.
Evaluation uses two splits from the same benchmark framework.
The paper’s retrieval experiments use a compact, controlled guideline corpus.
Evaluation relies primarily on LLM judges and a small targeted expert audit.
The model is designed for single-turn QA and does not represent longitudinal therapeutic context.

The model may generate incorrect, overly clinical, directive, or otherwise inappropriate content. Human review is required for any safety-sensitive use.

Ethical and Safety Considerations

This release is for research and reproducibility. It does not replace qualified mental-health professionals. Do not rely on the model for emergencies, diagnosis, medication decisions, or individualized treatment. Applications should implement independent safety checks, local crisis-resource handling, privacy protection, and qualified human oversight.

Code

Implementation and evaluation code:

https://github.com/jordy9090/selective-mental-health-rag

Citation

bibtex

@inproceedings{oh2026when,
  title     = {When Retrieval Helps: Selective Retrieval for Single-Turn Mental-Health QA},
  author    = {Oh, Hyunseo and Kim, Chong-Kwon and Choi, Yoonhyuk},
  booktitle = {KDD Undergraduate Consortium},
  year      = {2026}
}