The Model Hits Hugging Face With a 4-bit LoRA
A developer going by the handle 'ADHDAI' has released NeuroBait, a fine-tuned version of Mistral 7B trained specifically to generate text that triggers dopamine spikes in ADHD brains. The model was fine-tuned using QLoRA on a custom dataset of 12,000 curated snippets drawn from Reddit ADHD communities, productivity apps, and short-form fiction designed for rapid reward cycles. The resulting model is 4-bit quantized and fits in 4GB of VRAM. The developer claims it can produce 'micro-reward' responses that keep ADHD users engaged without the typical overstimulation or boredom. The model is available under an MIT license, allowing commercial use.
From LLM Burnout to Neurodiversity-Tuned AI
This isn't the first attempt to fine-tune a language model for neurodivergent audiences—startups like Goblin Tools have released ADHD-specific task-breakdown models—but NeuroBait is unique in targeting dopamine mechanics directly. Most LLMs are calibrated for neurotypical users: they prioritize coherence, politeness, and thoroughness. For ADHD brains, that often translates to boredom or overwhelm. The training data here is built around quick wins, variable rewards, and 'just one more' triggers, mimicking the pattern of TikTok feeds but with readable, substantive text. It's a cynical yet clever design choice.
The Real Win Isn't the Model—It's the Open Dataset
Honestly, the most interesting part isn't NeuroBait itself but the training recipe the developer open-sourced. If you're an ADHD researcher building tools for focus, having a curated dataset of dopamine-inducing prompts is far more useful than a black-box model. That said, the model could become a double-edged sword. It might help ADHD users stay on task when writing or browsing, but the same 'dopamine hooks' could easily be repurposed for addictive social media feeds. The developer explicitly warns against using it for adtech. But once the dataset is out, you can't police usage.
Where's the Evidence? No Clinical Validation Yet
NeuroBait is a prototype, not a treatment. The developer didn't conduct any controlled trials—fine-tuning loss curves don't measure dopamine levels. The training dataset was scored using a proxy metric based on user engagement in ADHD forums, which is hardly rigorous. It's also unclear whether the model's outputs actually increase dopamine or just mimic the linguistic patterns that accompany it. Without EEG or real-time feedback, we're in placebo territory. The Hugging Face model card has a 'limitations' section that reads like a liability waiver. Watch for replication attempts by academic labs before taking this seriously.