Gemma-2b-it-Psych-Merged is the full-weight, standalone version of the google/gemma-2b-it model, domain-adapted for psychological contexts. This model integrates the LoRA adapter weights from ecorbari/Gemma-2b-it-Psych directly into the base model using the merge_and_unload() process.

The model is optimized to generate empathetic, supportive, and professionally aligned psychological responses. Unlike the adapter-only version, this repository contains the complete merged weights, meaning it does not require the peft library for standard inference.

Model Details

Model Description

Author: Ederson Corbari (e@NeuroQuest.ai)
Date: February 01, 2026
Model type: Causal Language Model (LLM)
Language(s): English
License: MIT (consistent with Gemma base model terms)
Finetuned from model: google/gemma-2b-it
Merge method: LoRA Weights Integration (merge_and_unload)

Model Sources

Hugging Face Merged: ecorbari/Gemma-2b-it-Psych-Merged
Hugging Face LoRA Adapter: ecorbari/Gemma-2b-it-Psych
Base Model: google/gemma-2b-it

Uses

Direct Use

This merged model is ready for production and simplified inference. It can be loaded directly using standard transformers pipelines. It is intended for:

Research on empathetic AI behavior.
Educational demonstrations of domain-adapted LLMs.
Proof-of-concept psychological support tools.

Out-of-Scope Use

Clinical Use: This model is NOT a substitute for licensed mental health professionals. It must not be used for diagnosis or treatment.
High-Stakes Decision Making: It should not be used in autonomous counseling systems without human oversight.

How to Get Started with the Model

Since the weights are already merged, you can run inference using a simple pipeline:

python

import torch
from transformers import pipeline

model_id = "ecorbari/Gemma-2b-it-Psych-Merged"

pipe = pipeline(
    "text-generation",
    model=model_id,
    dtype=torch.float16,
    device_map="auto",
)

prompt = "I feel anxious and overwhelmed lately. What should I do?"
result = pipe(prompt, max_new_tokens=200)

print(result[0]["generated_text"])

Bias, Risks, and Limitations

Safety Disclaimer: The model may generate inaccurate information. Empathy in text generation does not imply clinical safety or medical correctness.

Data Bias: Responses may reflect biases inherent in the jkhedri/psychology-dataset.

Human Oversight: Users should apply human judgment, especially in sensitive conversational settings.

Training and Merge Process

The workflow involved loading the google/gemma-2b-it model in float16 precision, attaching the LoRA adapters trained on preference-based psychological data, and merging the weights into a single model for downstream use. This ensures compatibility with environments that do not support PEFT or require lower latency for inference.