一个 Backyard AI 项目。在家用小模型构建，带来大影响
什么是 ADHD？
ADHD（注意力缺陷/多动障碍）是一种常见的神经发育障碍，它影响大脑处理信息、调节注意力和控制冲动的方式。
一切始于我的妻子
我妻子患有 ADHD。我曾看着她站在一堆待洗衣物前，清楚知道该做什么，却依然无法开始。是卡住了，不是懒惰。对 ADHD 大脑来说，问题从来不是知道该做什么，而是从知道到开始之间的鸿沟。
大多数 ADHD AI 工具都搞错了这一点。它们专注于诊断和理论：测验、清单、整洁的待办事项列表。但对一个正陷入卡顿的人来说，待办清单只是多出来需要选择的东西，外加一个微弱的声音说"再努力一点"。正确，却完全没用。
于是我尝试了一些不同的做法。我把理论（执行功能障碍、任务启动瘫痪、兴趣驱动神经系统——可以说是 DSM 临床框架与《分心女王》的结合）拿去验证一个研究中很少有的东西：对我朝夕相处的人进行真实观察和实践。真正能让她解冻的是什么？不是计划。是火花。
NeuroBait 做了什么
NeuroBait 不会列待办清单。它会 点亮多巴胺，让"开始行动"变得有可能。
当你陷入停滞时，它会从你的对话中读取真正重要的东西——一个真实的截止日期、一个你在乎的人或事——然后用 3 到 6 句温暖自然的话语来回应（不用冷冰冰的标签，也不用一堵弹墙式的项目符号）：它理解你为什么卡住了，把你重新连接到你所热爱的东西上，并递给你 一个你现在就能做的小行动。"从那堆衣服的最上面拿一件下来。就一件。"
没有内疚。没有说教。你是主动行动的英雄，而不是病人。
技术栈
一个真实训练的模型，用了真实的预算：
- 基座：
google/gemma-3-12b-it（稠密的 Gemma 3 12B，标准Gemma3ForConditionalGeneration，因其能可靠地部署 transformers + peft 而被选中） - 方法：通过 Unsloth 使用 16-bit LoRA，而非 QLoRA
- LoRA：
r=16、alpha=16、dropout=0 - Epochs：
3| LR：2e-4| Batch：1 x grad_accum 8| 最大序列长度：2048 - 聊天模板：
gemma-3、回复标记<start_of_turn>user\n和<start_of_turn>model\n save_strategy="no"以规避已知的 Unsloth/TRL 检查点 pickle bug- 在 modal.com 上使用 H100 80GB GPU 训练
- 数据：规模小、人工精选、合成生成，源自真实的 ADHD 痛点，而非泛泛的效率套话。经验教训：对于语音而言，数据集质量胜过模型规模。
- 部署：在 ZeroGPU（zero-a10g）上的 Hugging Face Space，使用 Gradio + 标准 transformers + peft；基座模型以 4-bit bitsandbytes NF4 加载，并在运行时应用 LoRA 适配器。运行时不使用 Unsloth，也没有 GGUF 部署路径。
默认模型 vs. 微调模型
基座模型并不笨，它是有能力的。但开箱即用时，它仍会给你写一份待办清单，只不过是一份听起来更有共情心的待办清单：加粗标题、项目符号、泄露的标签、冗长的段落。对于一个"僵冻的大脑"来说，这堵"帮助"之墙本身就让人不堪重负。
这次微调在本质上表现出了不同的行为。它摒弃了那种结构，用温暖流畅的散文式语言说话。回复变短了。它先询问、再做假设。它会把你的上下文回扣给你，让回复感觉是写给你的，而不是写给一个笼统的"不堪重负的用户"。它学到的是语气，而不是台词。
不只适用于 ADHD
令我惊讶的是：这不仅仅是一个 ADHD 工具。
如今我们所有人都会撞上那堵墙：刷到大脑成一团浆糊，被臃肿的信息流淹没，连最简单的事情都无从下手。NeuroBait 的设计本身就是温暖的。对于任何在不堪重负中溺水的人来说，它都能起到多巴胺放松的作用，是一种温柔的、人性化的轻推，把你推回到一件小小的、可行的事情上。只不过 ADHD 的大脑最需要它。
未来的方向
它始于我家后院，只为了一个人。我们希望把它扩展到每一位 ADHD 患者，以及每一个曾感到卡住的人。
接下来：开放权重和完整流水线，支持双语文言（印尼语和英语），最重要的是，要与社区一起共建，而不是替社区构建。ADHD 工具长期以来一直由没有 ADHD 的人设计。真实的场景、真实的反应、真实的反馈，才是这个项目的核心。
如果你有 ADHD，或者你爱的人有，又或者只是经常感到不堪重负，欢迎来试用并告诉我哪里让你觉得烦。这些反馈才是全部意义所在。🧠⚡
A Backyard AI project. Built at home with a small model for a big impact
What is ADHD?
ADHD (Attention-Deficit/Hyperactivity Disorder) is a common neurodevelopmental that affects how the brain processes information, regulates attention, and controls impulses.
It started with my wife
My wife has ADHD. I've watched her stand in front of a pile of laundry, know exactly what to do, and still not be able to start. Frozen, not lazy. For an ADHD brain the problem was never knowing what to do. It's the gap between knowing and starting.
Most ADHD AI tools get this wrong. They focus on diagnosis and theory: quizzes, checklists, neat to-do lists. But a to-do list for someone in a freeze is just more to choose between, plus a faint voice saying "try harder." Correct, and completely useless.
So I tried something different. I took the theory (executive dysfunction, task-initiation paralysis, the interest-based nervous system, think DSM clinical framing meets The Queen of Distraction) and tested it against something the research rarely has: real observation and practice, on the person I live with. What actually unfreezes her? Not a plan. A spark.
What NeuroBait does
NeuroBait doesn't make to-do lists. It lights up dopamine to make starting feel possible.
When you're stuck, it reads your conversation for what matters, a real deadline, a person or thing you care about, and answers in 3 to 6 warm sentences that flow naturally (no clinical labels, no bullet walls): it gets why you're stuck, reconnects you to something you love, and hands you one tiny action you can do right now. "Pull one shirt off the top of the pile. Just one."
No guilt. No lecturing. You, the active hero, not the patient.
The stack
A real model, trained on a real budget:
- Base:
google/gemma-3-12b-it(dense Gemma 3 12B, standardGemma3ForConditionalGeneration, chosen for reliable transformers + peft deployment) - Method: 16-bit LoRA, not QLoRA, via Unsloth
- LoRA:
r=16,alpha=16,dropout=0 - Epochs:
3| LR:2e-4| Batch:1 x grad_accum 8| Max sequence:2048 - Chat template:
gemma-3, response markers<start_of_turn>user\nand<start_of_turn>model\n save_strategy="no"to avoid the known Unsloth/TRL checkpoint pickle bug- Trained on modal.com with H100 80GB GPU
- Data: small, hand-curated, synthetic, built from real ADHD friction, not generic productivity tropes. Lesson learned: for a voice, dataset quality beats model size.
- Deploy: Hugging Face Space on ZeroGPU (zero-a10g) using Gradio + standard transformers + peft; base loads in 4-bit bitsandbytes NF4 and applies the LoRA adapter at runtime. No Unsloth at runtime, no GGUF deployment path.
Default model vs. fine-tuned
The base model isn't dumb, it's capable. But out of the box it still writes you a to-do list, just a more empathetic-sounding one: bold headers, bullets, leaked labels, long paragraphs. For a frozen brain, that wall of "help" is itself overwhelming.
The fine-tune behaves differently in kind. It drops the structure and speaks in warm, flowing prose. It gets shorter. It asks before it assumes. It threads your context back to you, so it feels written for you, not for a generic "overwhelmed user." It learned the voice, not the script.
Not just for ADHD
Here's what surprised me: this isn't only an ADHD tool.
We all hit that wall now: doom-scrolled into mush, overwhelmed by bloated feeds, unable to start the simplest thing. NeuroBait is warm by design. It works as dopamine relaxation for anyone drowning in overwhelm, a gentle, human nudge back to one small, doable thing. ADHD brains just need it most.
Where it's going
It started in my backyard, for one person. The hope is to extend it to everyone with ADHD, and to anyone who's ever felt stuck.
Next: open weights plus the full pipeline, bilingual (Indonesian and English), and most important, built with the community, not for it. ADHD tools have too long been designed by people who don't have ADHD. Real scenarios, real reactions, real feedback are the project.
If you have ADHD, love someone who does, or just feel overwhelmed too often, come try it and tell me where it's annoying. That feedback is the whole point. 🧠⚡