Type a sentence in English or Swedish. Two versions of the Jepela engine judge it at the same moment: the standard engine,
and the same engine trained once on 324 example sentences with Jepela's fine-tuning service. Each answers with a probability for
question, statement and command.
Try:
The standard engine jepela
–
Trained on 324 examples
–
The exact question both are asked
How it was made
Trained once, not for every visitor. 324 labelled sentences in English and Swedish (157 questions, 109 statements and 58 commands)
went to Jepela's fine-tuning service, which trained this model in one run on the GPU. Every visitor uses that one model, and nothing you
type here is used to train it.
Kept only because it won. The service keeps a trained model only when it answers better than the standard engine on examples it
never trained on, and still listens to memory. On the 53 examples it held back: 51% right before training, 92% after.
What counts as a question. The writer asks something and expects an answer, also indirectly: "I wonder where the station is" and
"Jag undrar var stationen ligger" are questions. "She asked where the station was" reports a question, so it is a statement.
Only this demo uses it. The trained model belongs to the demo's own account. Everyone else who calls the API gets the standard
engine, unless they train a model of their own.
Measured on 80 other sentences it never trained on (57 English, 23 Swedish, with questions without a "?", statements with
question words, indirect questions and commands among them), through the live API on 2026-09-28: the standard engine answered 54 right
(67.5%), the trained model all 80. Those 80 also guided the training: examples were added round by round for the kinds of sentence
it still missed among them. So the fairer figure is on 73 new sentences, not among the training examples, written before the trained
model was asked (two sets added up): it got 69 right, the standard engine 47; most of its misses were Swedish exclamations such as
"Vilken fin dag!". The training took 38 seconds.