TinkerOne generation trained on this recipe: 42 documents.
temperature 1, sieve off, 0% real · loss 4.56 · rare words guessed 19.3%
Chapter 7 of 7
Workbench: a data recipe
The team must pick one recipe for the next model: a temperature, a sieve and a share of real data. Each was trained offline.
This chapter: Which recipe stays closest to real text and still guesses rare words?
About this chapter
You will learn
- How each choice moves loss and rare words
- Why real data matters most
- Why more than one recipe can be best
Words you’ll meet
- Recipe
- A temperature, a sieve and a share of real data.
- Real share
- How much of the training text is Larkspur’s own documents.
- Workbench
- The chapter’s free-play bench: change the recipe and see every result.
Step 1 of 2
The first is copies only. A recipe wins when none other has lower loss and more rare words guessed.
Change it on the .
Loss 4.56 · rare words guessed 19.3%BEATEN
| Recipe | temperature 1, sieve off, 0% real |
|---|---|
| Training documents | 42: 42 model-written, 0 real |
| Held-out loss | 4.56 (Larkspur mini: 4.04) |
| Rare words guessed | 19.3% (Larkspur mini: 27.1%) |
| Beaten by | temperature 0.7, sieve off, 50% real: loss 4.37, rare 23.5% |
Larkspur’s toy: Larkspur mini and copies of it, trained offline on Larkspur’s documents. It shows the mechanism, not a production model’s numbers.