Radical Numerics’ Tessera sorts training data by what it teaches a model

Radical Numerics says its Tessera method groups training documents by their measured effect on model performance, rather than by broad topics or where the data came from. In the company’s tests, this approach delivered four times the gain of mixing data across a hand-built topic taxonomy, offering a more fine-grained way to choose what models learn from.

Radical Numerics Watch analysis

What happened

Tessera first trains small proxy models on a sample of documents and measures how each affects selected evaluations. It uses those measurements to predict effects across the wider corpus, then groups documents with similar effects into thousands of derived sources. Instead of returning one set of source weights, the method fits tens of thousands of mixtures to map a Pareto frontier across evaluations. Developers can then choose a mixture that matches their desired balance of results.

Radical Numerics says the fourfold gain was measured against uniform sampling across six evaluations, compared with an optimised mixture using hand-defined topic categories. A random partition into the same number of sources produced no improvement, according to the company. It also applied the pipeline to DNA sequences, grouping them by their measured effect on genome-model evaluations rather than relying on biological categories.

Why it matters

Training data is not just a pile of examples: the proportions in which a model sees different material can affect what it learns. But broad buckets such as “science” or “books” may combine documents that help a chosen task with others that do not. Tessera’s approach aims to separate them by observed training effects, giving researchers more control over those trade-offs.

The practical prize is not a universal “best” mixture. The method maps alternatives across evaluations, so teams can select a balance that suits their aims rather than start again whenever priorities change. That could make data selection more targeted for language models and, if the reported DNA results hold up, for genome models too.

Our read

This is a substantial research idea because it turns data curation into something more measurable than sorting by topic and hoping for the best. The reported fourfold gain is promising, but it belongs to Radical Numerics’ own experiments, not a general guarantee that Tessera will improve every model or evaluation. The useful next step is replication on different datasets and tasks. At least the company has offered a method readers can scrutinise, rather than asking us to applaud the word “optimisation” and move along.

What to watch

• Whether independent evaluations reproduce the reported gains beyond the six tests described.
• How the method performs as the corpus, model size and training budget grow.
• Whether the DNA application yields useful results on evaluations beyond those in the company’s post.

Discussion spark: Should model builders optimise training data for a chosen set of evaluations, or does that risk making models better at the tests than at the work they are meant to do?

not affiliated with or endorsed by Radical Numerics