NeuroBait: I fine-tuned a model to spark dopamine for ADHD brain
A fine-tuned Gemma 3 model called NeuroBait tries to help ADHD overwhelm by turning context into one tiny, warm next step instead of a task list.
Intelligence analysis by GPT-5.4 Mini
Built for the author’s wife, who has ADHD, NeuroBait targets task-initiation paralysis: it reads what matters in a conversation, then replies in warm prose that makes starting feel possible. The project fine-tunes Gemma 3 12B with LoRA on hand-curated ADHD scenarios and deploys it on a Hugging Face Space.
NeuroBait is like a kind friend who stands beside a kid staring at a messy room and says, “Pick up just one shirt.” It does not shout a big list; it gives one small step so starting feels less scary.
Analysis
What it is
NeuroBait is a custom model meant to help with ADHD task-starting paralysis, not to generate another checklist. The author says the core problem is the gap between knowing what needs to be done and being able to begin, so the model aims to create a small spark rather than more instructions.
How it behaves
Instead of bullet points or clinical framing, it reads the conversation for what matters most, such as a deadline or a person the user cares about. It then answers in three to six warm sentences, reconnects the person to something meaningful, and offers one tiny action that feels doable right now.
How it was built
The system starts from google/gemma-3-12b-it and was fine-tuned with 16-bit LoRA using Unsloth. The setup uses r=16, alpha=16, zero dropout, three epochs, a 2e-4 learning rate, batch size 1 with gradient accumulation of 8, and a 2048 token sequence length. Training ran on Modal with an H100 80GB GPU. The dataset was described as small, hand-curated, and synthetic, but based on real ADHD friction rather than generic productivity advice.
Deployment and direction
The model is deployed as a Hugging Face Space on ZeroGPU with Gradio, transformers, and peft. At runtime, the base model loads in 4-bit NF4 and the LoRA adapter is applied on top. The author says the fine-tune learned the voice, not just the script, and that the project may expand to open weights, bilingual Indonesian and English support, and community feedback.
Key points
- NeuroBait is designed to help with task-initiation paralysis, not to produce traditional to-do lists.
- The author built it after observing how ADHD can make starting a task feel impossible even when the next step is obvious.
- The model fine-tunes Gemma 3 12B with LoRA and was trained on a small, hand-curated dataset based on real ADHD friction.
- Its output is meant to be warm, brief, and action-oriented, with one tiny step instead of a wall of instructions.
- The project is deployed on a Hugging Face Space and may later add open weights and bilingual support.
If the project keeps its focus on real ADHD situations, it could become a more useful support tool than generic task apps. The planned open weights, bilingual support, and community feedback could make it easier for more people to try and improve.
The model is still built from a small, hand-curated dataset, so it may not generalize well to every kind of overwhelm or ADHD experience. If it misses the user’s context, it could drift back toward the same kind of generic help the article says is already unhelpful.
