If it's not in front of me, it's gone: I fine-tuned a 4B model for my friend with ADHD

Description and use

The "Actually" app keeps ADHD users’ tasks visible and predicts realistic durations using a fine‑tuned 4B model and TabPFN, running locally on a single laptop.

Problem

A friend with ADHD forgets tasks that are not physically present and gets hyperfocused on one task, causing other tasks to be ignored.

Outcome

The fine‑tuned 4B model achieved 100% valid JSON and 97% task F1 on held‑out dumps, outperforming zero‑shot baselines. The app runs locally on a single laptop and costs $2.81 to build.

Workflow

• 1. User types a brain‑dump in natural language.
• 2. A 4B Qwen3.5 model fine‑tuned on Tinker parses the dump into structured tasks in JSON.
• 3. TabPFN reads the user’s past guessed vs actual durations and predicts a realistic duration, confidence range, and finish‑today probability for each task.
• 4. The UI displays tasks with estimated durations, ranges, and a "Move to tomorrow" flag for tasks that don’t fit the day.
• 5. User marks tasks as done to remove them from the list.

Evidence

If infront remember or else gone from brain. person or task

Source

Qwen3.5-4B + LoRA 100% valid JSON, 97% task F1

Source

The parser: a 4B model that reads a brain‑dump.

Source