Making the unstartable, startable: why your AI assistant needs a dopamine nudge
For creators, artists, and anyone wrestling with a mental block, the current generation of productivity software often feels like shouting into a void. Most tools assume the problem is a lack of information or organisation. They dump you with checklists, bold headers, and clinical frameworks that only add to the paralysis. This approach fails because it misunderstands the core issue: the gap between knowing what to do and actually initiating the action.
NeuroBait flips the script. It is not a task manager. It is a fine-tuned model designed to bridge the gap between intention and execution by triggering a micro-dose of dopamine. Instead of lecturing the user on why they should be productive, it identifies a specific emotional hook—whether that is a loved one, a creative project, or a looming deadline—and responds with three to six warm, flowing sentences. The output avoids clinical labels and bullet points, offering a single, tiny action to take immediately. “Pull one shirt off the top of the pile. Just one.”
The system is built on a specific technical stack chosen for reliability and efficiency:
- Base Model:
google/gemma-3-12b-it
, a dense Gemma 3 12B model selected for stable transformers and PEFT deployment capabilities.
- Training Method: 16-bit LoRA via Unsloth, avoiding the complexity of QLoRA.
- Hyperparameters: Rank
r=16
, alpha
alpha=16
, with a dropout rate of
0
.
- Training Schedule:
3
epochs, learning rate
2e-4
, batch size
1 x grad_accum 8
, and a maximum sequence length of
2048
.
- Chat Template:
gemma-3
using response markers
<start_of_turn>user\n
and
<start_of_turn>model\n
.
- Checkpointing Strategy:
save_strategy="no"
to prevent known Unsloth/TRL checkpoint pickle bugs.
- Compute Environment: Trained on Modal.com using H100 80GB GPUs.
- Data Source: A small, hand-curated synthetic dataset derived from real observations of ADHD friction rather than generic productivity tropes.
- Deployment: Hosted on a Hugging Face Space using ZeroGPU (zero-a10g) via Gradio and standard transformers. The base model loads in 4-bit bitsandbytes NF4, applying the LoRA adapter at runtime. No Unsloth or GGUF paths are used in production.
The distinction between the base model and the fine-tuned version is stark. Out of the box, the base model is capable but fails to understand the user’s state. It generates empathetic-sounding to-do lists with bold headers and leaked labels, creating a wall of text that overwhelms a frozen mind. The fine-tuned model behaves differently; it drops the rigid structure entirely. It speaks in warm, flowing prose that feels personally addressed. It asks before assuming and threads the user’s specific context back to them, ensuring the response feels written for an individual rather than a generic “overwhelmed user.”
While the tool was born from personal observation of a spouse with ADHD, its utility extends beyond that specific demographic. The creator notes that anyone who has ever doom-scrolled into a state of mush or been unable to start a simple task can benefit from this “dopamine relaxation.” It acts as a gentle, human nudge back to one small, doable thing. The project aims to release open weights and a full pipeline, including bilingual support for Indonesian and English, built with the community rather than just for them.
The current build is available for testing at the Hugging Face Space link below. The creator explicitly invites feedback from users with ADHD, their loved ones, or anyone who feels overwhelmed too often. The project is defined by real scenarios, real reactions, and real feedback.
Try the tool here: https://huggingface.co/spaces/build-small-hackathon/NeuroBait
Key takeaways
- NeuroBait is not a productivity app; it is a fine-tuned model designed to spark the dopamine required to initiate action.
- The model was trained on 3 epochs using 16-bit LoRA on H100 GPUs, specifically avoiding generic productivity tropes in favour of real-world friction data.
- Deployment runs on a zero-a10g instance using Gradio, loading a 4-bit base model and applying the LoRA adapter at runtime.
- The tool is currently open for community feedback and plans to expand with bilingual support and open weights.
Stay ahead of AI. Get the most important stories delivered to your inbox — no spam, no noise.
