Setup
A 2.1B open model was fine-tuned on 115,580 sub-task-discovery problems drawn from 259 system-produced plans — about one GPU-day of supervised training. The reference is its 27B teacher. Held-out evaluations were blind-scored by an independent frontier model from a different provider, on tasks from plans the student never trained on.
Result
On fine-grained planning work the tuned 2B reached 94% of its 27B teacher — 18 of 18 cases, blind-judged. The untuned base model falls far behind on the same tasks: the skill demonstrably came from the generated plans.
The more surprising measurement is what the tuned model already holds. Sampled eight times per task with the best answer kept, it equals the teacher outright — and roughly a quarter of its single attempts match or beat the teacher's answer for the same task. The remaining gap is not a ceiling on capability; it is a matter of consistency — which is precisely the kind of gap that further preference training works on, using only the model's own attempts, ranked.
Why it matters
Specialists only make economic sense if capability can be moved into small, cheap models. A 13× size gap closed to 94% on a narrow skill, using only generated data, is the first direct evidence that our corpus transfers into weights at all — the necessary first step before the full-scale claim, which remains to be proven.
This establishes
- A first transfer into weights: one planning skill, learned from generated plans alone, blind-judged at 94% of a 13×-larger teacher.
- The capability ceiling of the tuned 2B is at the teacher's level — best-of-eight equals it; the residual gap is consistency.
companion technical note — methods, limitations, judge protocol — available on request · hello@truna.ai