NeuroBait: The AI That Stimulates Dopamine for ADHD

NeuroBait: A Revolution for ADHD

Attention Deficit Hyperactivity Disorder (ADHD) is a neurodevelopmental disorder that affects how the brain processes information, regulates attention, and controls impulses. This common disorder presents significant challenges, particularly in the transition from knowledge to action, a problem often poorly addressed by traditional tools.

An Approach Inspired by Personal Experience

The story of NeuroBait begins with a personal observation. The creator, faced with the task initiation paralysis of his wife with ADHD, sought to understand what could truly unlock this situation. He found that most AI tools for ADHD focus on diagnosis and theory, offering quizzes, checklists, and neatly organized tasks. However, for a frozen brain, these lists are just another source of choices, often unhelpful.

An Innovative Approach: Stimulating Dopamine

NeuroBait stands out by not creating task lists. Instead, it stimulates dopamine to make starting possible. When a person is stuck, NeuroBait analyzes the conversation to identify what truly matters to them: a deadline, a person, or something dear to their heart. It responds with 3 to 6 warm phrases that flow naturally, without clinical labels or bullet points. The goal is to reconnect the user to something they love and suggest a small, achievable action immediately.

Technical Details and Fine-Tuning Method

The base model used is google/gemma-3-12b-it, chosen for its reliable transformers and efficient deployment. Fine-tuning was performed using the 16-bit LoRA method, with specific specifications: r=16, alpha=16, dropout=0. Training parameters include 3 epochs, a learning rate (LR) of 2e-4, a batch size of 1 with a gradient accumulation of 8, and a maximum sequence length of 2048. The model loads in 4 bits using bitsandbytes NF4, and the LoRA adapter is applied in real-time. Training was conducted on modal.com with an H100 80GB GPU. The data used is small, carefully selected, and synthetic, built from the real friction associated with ADHD, avoiding generic productivity tropes.

Deployment on Hugging Face Space

NeuroBait is deployed on Hugging Face Space, using ZeroGPU (zero-a10g) with Gradio and standard transformers. This deployment allows for a smooth user experience, without the need for Unsloth in real-time. The default model, while empathetic, tends to produce task lists with bold headers and long paragraphs, which can be overwhelming for a frozen brain. The fine-tuning of the model abandons this rigid structure to express itself in warm, flowing prose, asking questions before making assumptions. The chat template used is gemma-3, with response markers <start_of_turn>user\n and <start_of_turn>model\n. The "save_strategy" backup strategy is set to "no" to avoid the known Unsloth/TRL checkpoint pickle bug.

A Tool for Everyone, Not Just for ADHD

Although initially designed for ADHD, NeuroBait proves useful for anyone feeling overwhelmed by tasks or incessant information flows. Its warm and human design aims to provide gentle and achievable support. The project aspires to expand to a broader community, including users from different languages and cultures, with a goal of bilingual development in Indonesian and English.

Towards a Collaborative Future

The goal of NeuroBait is to expand to a larger community, including users from various languages and cultures. The project aims to be built with the community, integrating real feedback to continuously improve the tool. Upcoming steps include opening the weights and the complete pipeline, and above all, building with the community, not just for it. Tools for ADHD have long been designed by people who do not have ADHD. Real scenarios, real reactions, real feedback are the project. If you have ADHD, love someone who does, or just feel overwhelmed too often, come try it out and share your thoughts.