NeuroBait: I fine-tuned a model to spark dopamine for ADHD brain
Inspired by his wife's ADHD struggles, the author fine-tuned NeuroBait to replace to-do lists with warm, spark-like dialogue, proving that small, domain-specific data can create more empathetic AI assistants.
- The core ADHD barrier isn't a lack of plans, but a lack of a 'spark' — a tiny, pressure-free cue to start.
- NeuroBait uses 3-6 warm sentences to reconnect users with what matters and offers one immediate, minuscule action, avoiding diagnostic labels.
- Fine-tuning Gemma-3-12B with 16-bit LoRA on a hand-curated small dataset proves that for shaping a 'voice,' data quality far outweighs model size.
- This highlights a trend in AI mental health: shifting from generic assistants to personalized, small-data fine-tuning, and from management tools to emotional catalysts.
The Trigger: The Wife Standing Before the Laundry Pile
Most ADHD tools assume you need more to-do lists, reminders, and check-ins. But the author watched his wife stand in front of a pile of laundry, knowing exactly what to do yet unable to start—not lazy, but gripped by "task initiation paralysis." Existing AI assistants still spit out to-do lists, even if wrapped in empathy, which become a wall of text that suffocates rather than helps. So he decided to fine-tune a model that doesn't make plans; it creates a spark.
How NeuroBait Works
It doesn't diagnose, lecture, or theorize. When you're stuck in a conversation, it reads for what truly matters to you—a real deadline, a person or thing you care about—and responds in 3 to 6 warm, natural sentences. It might say, "Pull one shirt off the top of the pile. Just one." No guilt, no preaching. It treats you as the active hero, not the patient.
Under the hood: a fine-tuned google/gemma-3-12b-it using 16-bit LoRA (not QLoRA) with rank 16, alpha 16, trained for 3 epochs on a small, hand-curated synthetic dataset built from real ADHD friction points. The key lesson: for shaping a "voice," dataset quality far exceeds model size. Deployed on a Hugging Face Space with 4-bit quantization and LoRA adapter applied at runtime, it runs cheaply.
Trend Insight: The Next Paradigm in AI Mental Health
This project highlights a broader shift: AI in mental health is moving from "one-size-fits-all brains" to "specialized souls." While we once chased bigger models, a small model that understands a specific pain point—like task initiation in ADHD—can be far more effective than a general-purpose giant. More importantly, the AI's role transforms from a management tool into an emotional trigger. It doesn't just organize tasks; it sparks dopamine so action feels possible. This reflects a rethinking of behavior change: sometimes, you don't need to know more; you just need to feel capable of starting.
Practical Takeaways for Developers, Users, and the ADHD Community
If you're building AI for a niche audience—whether neurodivergent support, mental health, or any context requiring deep empathy—don't rely solely on large model zero-shot. Observe real friction, gather a small high-quality dataset, and fine-tune a model with a warm, attuned voice. LoRA makes such experiments affordable. For those with ADHD, tools like NeuroBait remind us that overcoming procrastination isn't about willpower but about designing a laughably low barrier to entry. For the AI industry, it proves that emotional precision can stem from small data, as long as you truly understand the user's "flow."
Counterintuitive Take: Why To-Do Lists Are the Worst Solution
You'd think giving a brain clear tasks is helpful, but for ADHD, options are burdens. Neuroscience tells us that executive dysfunction makes decision-making extremely energy-draining, so more structure can lead to paralysis. NeuroBait flips the script: it eliminates choice entirely and prescribes one tiny, concrete, even seemingly insignificant action. This actually releases dopamine because the brain anticipates, "I can do this." It's not a productivity tool; it's an action catalyst.
Built with love and keen observation, this tiny tool reveals the most touching side of AI: it's not about being smarter than humans, but about knowing when to stop being so clever.
Analysis by BitByAI · Read original