Houjun LabSmall experiments,
measured honestly.
Houjun Lab is where we write up research projects run on a single Mac: how language models learn from text they have never seen, what they forget on the way, and how to tell the difference. Each project fixes its tests before training, keeps the results that did not work, and links to code you can run again.
Projects
One page per project: the question, the design, what was measured and where it stands.
MedAdapt Lab
Can a small language model learn new findings by reading research papers? Five experiments on openly licensed 2026 ADHD papers and on fictional trials: what it learns, what it forgets, and why rewording beats repetition.
Read the project →More to come
Further studies will be added here as they are designed. Every project gets the same treatment: a fixed design, open code and every result reported.
How we work
The rules every project here follows, so that a result means what it says.
Tests before training
Exam questions and evaluation texts are written and frozen first. Their checksums are recorded, so a test cannot quietly change after the results are known.
A control for every claim
A gain only counts if a matched control does not show it too: material the model was trained on is always compared with material it was not.
Negative results stay
Runs that failed, effects that did not appear and measurement mistakes are reported next to the successes, with what we changed and why.
Reproducible on one machine
Pinned model revisions, hashed data and logged settings. Everything runs locally on an Apple Silicon Mac with open tools.
Licensed data only
We train only on text whose licence allows it, check that licence item by item, and do not redistribute anyone else's data or papers.
Research, not advice
These are experiments on how models behave. Nothing here is medical, legal or product advice, and no model from these projects is fit for real-world decisions.
Also from Houjun
Finished, open-source apps that run AI models locally on your Mac.