Braid
Training data with a clear source, clear usage rights, and repeatable test sets.
Machine-learning research · Richmond, BC
Learn what makes AI smarter. Then scale it.
Cenetex Lab trains small language models. Their size lets us see what helps: better data, better training order, or better tests.
Zero C3.3 · progress saved4.85M
settings learned by the same small research model
11.1% lowertest loss after changing only the training order
57.07% fewerpieces needed to store the same text
19.3Mpieces of text in the C3.3 training plan
Why we start small
Big models are powerful. Small models help us learn why.
Large models can hide what caused an improvement. Zero is small on purpose. We keep the model the same, change one thing, and test whether it really helped.
If an idea works, we can try it on a larger model. If it fails, we learn why before spending much more money.
The research system
Zero
The same small model used to test better data, better training order, and better ways to learn.
Grounded tests
Tests that check facts and relationships, not only whether the writing sounds good.
A clear improvement
−11.1%
Training order made a real difference.
In C3.1, we mixed the training tasks instead of teaching them in separate blocks. Test loss fell from 1.7623 to 1.5666. Everything else stayed the same.
Same modelSame dataSame computing time
Zero lineage
One model. One clear lesson at a time.
Make the tools reliable
We built a way to split text that loses nothing, used facts we can trace, and made training repeatable.
Start learning language
After reading 11.9 million pieces of text, the model began to show a basic sense of language and meaning.
Find a weak spot
The model improved, but often changed its answer when we reversed the order of two choices.
Teach both directions
We now show both orders in the same learning step. The final result is not ready yet.
Current experiment · C3.3
Show both orders in the same learning step.
The model now sees a pair and its reversed version at the same time. We saved its progress after 7,000 of 9,442 training steps. We will not call it a success until the final tests are complete.
7,000 complete9,442 planned
What must improve
+10 pointssame answer after reversal
+5 pointsboth versions correct
≤1%cost to normal language skill
Build with Cenetex
Bring useful data, computing power, or a hard ML problem.
We work with teams that want measurable machine learning—not a demo that only looks intelligent.
Talk to Cenetex →