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

Source: HuggingFace Date: 2026-06-09 URL: https://huggingface.co/blog/build-small-hackathon/neurobait-adhd

Summary

A hackathon entry from HuggingFace’s “Build Small” event: a Gemma 3 12B model fine-tuned with 16-bit LoRA (r=16, α=16) on hand-curated synthetic data targeting ADHD task-initiation paralysis. Rather than generating task lists, the fine-tuned model produces a single micro-action in warm conversational prose, distinguishing itself from the base model’s generic advice voice. The deployment stack runs on HuggingFace Spaces with 4-bit quantization via bitsandbytes — accessible on commodity hardware, no dedicated inference server required.

Implications

Feeds the local-model layer thread. A few points worth tracking:

• The technique confirms that small fine-tunes on Gemma 3 12B can shift voice and interaction style substantially without retraining the full stack — the model didn’t gain new facts, it gained a persona. That’s the pattern to watch as practitioners reach for LoRA to build narrow assistants tuned to a user archetype rather than general capability.
• The hardware profile (H100 for training, bitsandbytes 4-bit for inference on Spaces) is illustrative of the current sweet spot: fine-tune once on rental compute, serve cheap. Relevant for any project building domain-specific assistants on open-weight models.
• Weak signal for the voices/power dynamics thread: fine-tuning for a specific neurotype’s attention pattern is an early form of model personalization that puts behavioural shaping in the hands of individual practitioners. Worth watching at scale.