Inspiration
The reasoning capabilities of a model is very good and we could use it in health care sector.
Fine-tuning GPT-OSS (20B) as a MindFullGPT — A Mental Well-Being Assistant Model trained with Unsloth
What it does
A model that can offer a supportive ear, guide users through simple grounding techniques, and provide helpful reframing exercises based on Cognitive Behavioral Therapy (CBT) principles.
How we built it
We build it with unsloth, LoRa adapters as well as by utilizing free GPU's provided by the kaggle, google colab
Challenges we ran into
Model Size & Compute Requirements-- GPT-OSS 20B is extremely large, and running it efficiently required significant GPU resources. We had to optimize with LoRA fine-tuning and careful batching to make training and inference feasible within hackathon constraints.
Fine-Tuning Data-- Finding appropriate datasets for mental health conversations was difficult. Many publicly available datasets are either too small, too clinical, or not suitable for empathetic, supportive dialogue. We had to carefully curate and align data for safe training.
Accomplishments that we're proud of
We have put huge effort into making a working model even with the hardware and software limitations as well as a service for people who might be struggling with mental health issues.
What we learned
Training a model is no easy task it requires a lot of trial and error and huge amounts of patience as well as some heavy duty equipment.
What's next for MindFull-GPT
Expand the prototype into a full mobile and multi-platform app, making mental health support accessible anytime. We plan to add multilingual capabilities, personalized features like journaling and mood tracking, and stronger crisis detection with direct links to hotlines and professional networks. To ensure wider reach, we’ll explore lightweight, optimized model versions for low-resource environments, while also collaborating with psychologists and NGOs to keep the system ethical, safe, and aligned with real-world therapeutic practices.
Built With
- kaggle
- python
Log in or sign up for Devpost to join the conversation.