Why
The PoC distills a small model from a larger one's outputs on a narrow task and measures the capability retained versus the size saved.
How it works
Not yet built.
Where it lands in Jayverse
- Number: distill any on-site prediction model from a larger reference model. If Number ships an indicator or prediction model, train it on a narrow task from a bigger teacher's outputs, and publish capability-retained versus size-saved next to the indicator.
- Verex: use a distilled model as the cheap first-pass resolution check. Keep a larger model or a human as the escalation path when the small model's triage disagrees with the expected outcome.
- Auditor: require the teacher, task and retained-capability note for any distilled model in production. A distilled model's output should never be judged as if it came from the teacher without that methodology written down.
Key expressions
| Expression | 뜻 · 쓰이는 자리 |
|---|---|
| at a fraction of the cost | 원래 비용의 일부만으로 · 훨씬 적은 비용으로 비슷한 효과를 낼 때. "at a fraction of the cost" |
| the standard path from X to Y | X에서 Y로 가는 표준 경로 · 일반적으로 거치는 정해진 과정을 말할 때. "the standard path from a frontier model" |
| retained versus saved | 유지된 것 대 절약된 것 · 성능 유지량과 비용 절감량을 대비할 때. "the capability retained versus the size saved" |
| narrow (task) | 범위가 좁은 작업 · 한정된 특정 작업만을 다룰 때. "on a narrow task" |
| capture (capability) | 능력을 포착해 담아내다 · 큰 모델의 성능을 작은 모델이 흡수할 때. "capturing much of the capability" |
| frontier model | 프런티어 모델(현재 최고 수준의 최첨단 대형 모델) · 증류의 출발점이 되는 고성능 대형 모델을 가리키는 업계 용어. "the standard path from a frontier model to a" |