Y0 Fine-tuning
previewYour distribution, not ours — adapters trained on your traces, owned by you.
[ 01 ]Spec sheet
Y0 Fine-tuning exists for the gap general models cannot close: your taxonomy, your tone, your edge cases. It trains lightweight adapters on top of Y0 Text and Y0 Embeddings rather than forking whole models, which keeps tuned variants cheap to serve, fast to iterate, and always current with base-model improvements — when the base family ships a better release, your adapter rides along after an automatic evaluation pass confirms no regression on your held-out set. The workflow is deliberately evaluation-first. You assemble a dataset (often directly from corrected run traces, which are already in the right format), the platform splits off a holdout, and every training job ends with a scored comparison against the untuned base on your own examples — if the adapter does not beat the base, the job tells you so instead of letting you ship a regression. Data boundaries are absolute: your training data trains your adapters only, adapters are project-scoped resources behind the same key scopes as everything else, and deleting a dataset deletes the adapters derived from it. In preview for Scale-tier customers; the waitlist is open and worked through weekly.
[ 02 ]Capabilities
Adapter training over Y0 Text and Y0 Embeddings bases
Dataset assembly directly from corrected run traces
Automatic holdout evaluation against the untuned base
Zero-downtime adapter promotion and instant rollback
Strict tenancy — your data trains your models, nothing else
[ 03 ]Best for
Domain taxonomies — claims codes, legal clause types, SKU classes
Brand and personal voice that prompting cannot hold steady
High-volume narrow tasks where tuned-fast beats general-deep
[ 04 ]Sample request
{
"base_model": "y0-text-fast",
"dataset": "ds_claims_v3",
"method": "adapter",
"eval": { "holdout": 0.1, "metric": "exact_match" },
"suffix": "claims-coder"
}