Cantara Labs: giving LLMs taste
Giving LLMs taste.
We align open-weight language models to what people with taste actually prefer, so a draft reads like someone chose to write it.
01
Correct, fluent,
and forgettable.
Ask a capable model for a post and it hedges, opens with a question nobody asked, and sounds like every other model. The training that made it helpful also taught it a safe, average register. Taste is beating average on purpose, and we break it into parts we can measure: specifics, voice, restraint and structure.
− default model
+ with taste
02
Five steps,
one loop.
Better drafts get more honest edits, and those edits become the next round of training data.
- 01
We build on open-weight base and instruct models, so we can change a model at the level of its parameters, measure what moved, and run it where our customers need it.
- 02
Editors and writers compare two drafts and pick the one they would publish, with a short reason. Pairwise choices are far more consistent than a score out of ten.
- 03
A writer's rewrite shows exactly what the model got wrong. With their permission, the draft becomes the rejected answer and the published version the chosen one.
- 04
Preference optimization (DPO and its relatives) and reward models move the defaults: shorter openers, concrete details, fewer filler phrases, a closer match to the author's voice.
- 05
Readers see two unlabeled drafts and pick the one they would rather read. A change ships only when it wins.
03
What we hold to.
- 01
Taste is the hard problem left
Proofs and code can be checked, so models got good at them. Good writing has no unit test, which is why most labs leave it alone.
- 02
Writers stay the authors
The model drafts and people decide. Nothing Cantara writes goes out under someone's name until they approve it.
- 03
Content writing first
We start where Cantara already works, posts drafted from real work. Then the same model goes to other verticals where writing decides the outcome.