NeuroBait: I fine-tuned a model to spark dopamine for ADHD brain

A Targeted Model for ADHD Task Initiation

Developer Harisabekti Dicky Subrata has released NeuroBait, a fine-tuned language model designed specifically to help individuals with ADHD overcome task-initiation paralysis. The project, hosted on Hugging Face Spaces, avoids conventional productivity aids like to-do lists, which can often increase overwhelm. Instead, NeuroBait analyzes a user's conversational context to provide a short, warm, and empathetic prompt—typically 3 to 6 sentences—that reconnects them with their motivation and suggests a single, small, actionable step. The goal is not to organize tasks, but to generate the 'dopamine spark' needed to bridge the gap between knowing what to do and actually starting.

The NeuroBait Stack

The project emphasizes effectiveness on a small budget, leveraging open-source tools and efficient cloud infrastructure. The developer notes that the quality of the small, hand-curated synthetic dataset—built from real-world scenarios rather than generic productivity tropes—was more critical to achieving the desired conversational voice than the size of the model itself. The technical foundation includes:

• Base Model: google/gemma-3-12b-it
• Fine-tuning Method: 16-bit LoRA (r=16, alpha=16) using Unsloth for optimization.
• Training Infrastructure: A single H100 80GB GPU on modal.com.
• Deployment: The model is deployed on a Hugging Face Space using a ZeroGPU instance, with the base model loaded in 4-bit and the LoRA adapter applied at runtime.

Impact on Specialized AI Development

NeuroBait serves as a practical example of how targeted fine-tuning can create a significantly different user experience compared to a base model. While the stock Gemma 3 model provides empathetic but structured responses (lists, bold headers), the fine-tuned version learns a specific 'voice' that is more fluid and less intimidating for a neurodivergent user. This approach has broader implications beyond ADHD, targeting the general sense of digital overwhelm many people experience. The project's roadmap, which includes open-sourcing the weights and building a bilingual model (Indonesian and English) with community feedback, signals a move toward more inclusive, user-centric AI tools developed by, and for, the communities they aim to serve.

NeuroBait's success demonstrates that for specialized AI applications requiring a specific voice and emotional tone, the value proposition is shifting from raw model scale to the authenticity and quality of the training data.