Research
How we build and measure Atom.
Model reports, experiments and engineering notes, labelled by kind. Each one states its sample sizes and its limitations.
Research
Research
Experiments and model reports: questions about how Atom learns, with the evidence for and against.
- Research
Does on-policy preference optimisation help after targeted fine-tuning?
Twice we added a round of preference optimisation after targeted fine-tuning. Once it made no measurable difference, and once it scored lower. Here is the evidence and an untested hypothesis for why.
Read - Research
Atom: model overview and evaluation report
What Atom is, how we evaluate it, the benchmark results we have for each release, and where it falls short.
Read
Engineering
Engineering
Release reports and evaluations of the systems we ship.
- Engineering
Coding tasks in our agent training harness: probe4 against probe5
On 40 generated coding tasks, the pass rate rose from 15% to 75% after agent training. What this measures, and the important things it does not.
Read - Engineering
Atom v1.2 release report
How Atom v1.2 was made, what we measured on our development set, what got worse, and what we have not measured yet.
Read
Product
Product
Launches and changes to Incbury's products.
- Product
Incbury Code now has a desktop app
Incbury Code, our coding agent, now comes as a desktop app for Windows and Linux alongside the terminal and the VS Code and Cursor sidebar.
Read - Product
Introducing Incbury and Atom
Incbury Chat is open to everyone, powered by Atom, the mixture-of-experts model we designed, trained and serve ourselves.
Read
Directions
Open questions.
What we are investigating next. These are questions, not results.
Open question
Targeted data over generic mixtures
On our development set, targeted, machine-checked training examples have helped Atom more than generic fine-tuning mixtures. We do not yet know how far that carries over to unseen tasks.
Open question
Tool use in preference training
Preference optimisation lowered Atom's development-set score both times we tried it. One hypothesis is that our sampler cannot run the Python Atom writes, so correct tool-using answers were marked as rejected.
Open question
Asking the right question
Atom now asks when a request is ambiguous, but its questions are often too vague to help.
Open question
Knowledge in a small model
Atom's general knowledge is weak: v1.1 scored 27.2% on MMLU. Search helps, but only when Atom recognises that it needs to look something up.
Open question
Measuring the agent, not just the model
Our coding-agent results so far come from a harness run on tasks related to our training data.