Glossary

Adapter (LoRA). Small extra weights attached to a model's layers. JuL's cross models on the preset's own weights are LoRA adapters, switched on only while a question type they read is answered.

autotune. Trains a head per question from labeled examples and keeps it only if it beats zero-shot in cross-validation. More.

Backend. What runs the model: mlx, torch, onnx or api. More.

Bundle. A directory written by jul pack: a fixed set of questions with everything that does not depend on the state precomputed. Holds no weights. More.

Calibration. Making probabilities mean what they say. JuL fits a temperature per model, and autotune a temperature and a bias per option on your labels.

Center. The vector subtracted before comparing the state with the options: set per preset (the mean of the options for most, a generic one for minicpm5-2b), the mean of a context's examples when it has some.

Choice. A question that picks one option among several. Answers choice, probabilities, confidence.

Confidence. For a Choice, the chosen option's probability. For a Score, how peaked the levels' distribution is (1 = one level takes everything). A Noul's is max(noul, 1 − noul) where JuL needs one (escalation).

Context. Unlabeled examples (and optionally a description) of your data, saved under a name; also where heads are stored. More.

Cross model. A reading that takes the question and the text together, in one pass, for the questions an embedding model answers near chance. More.

Decision bench. 2,108 typed questions over 12 task families, in English and French, used to compare JuL with Jev on Choice, Noul and Score. Results.

Decision model. A model trained to answer typed questions, which brings its own format in a decision.json (pointer reading). More.

Description. What the model reads for an option. The key is what you get back.

ECE (expected calibration error). How far the probabilities are from the observed accuracy; lower is better. The default model: 0.084, Jev: 0.156.

Encoder. A small bidirectional model (BERT, e5), read as its sentence embedding. Fast on CPU. More.

Escalation. Answering locally and sending only the questions below a confidence bar to the next decider. More.

Formulation. One prompt a state is read with, at one layer: one_word (the state alone, shared by every question), question_options, question. A preset uses two by default.

Head. A small classifier trained on the vectors (or on TF-IDF, or both) to correct a model on your task. The model itself is never changed.

Hidden state. The vector a model builds at a layer for a token. JuL reads it instead of letting the model write.

Hybrid head. A head that reads both the model's vectors and the TF-IDF of the text.

Input limit. The most tokens a reading takes (a cross.json's max_length, a decision.json's max_state_tokens, an encoder's positions). Past it the end is dropped, logged and counted in usage.truncated_tokens; on_long="error" refuses the call instead. More.

Jev. TypeSafe's hosted decision model, whose SDK and HTTP protocol JuL follows. JuL is not affiliated with TypeSafe.

Jev bench. Jev's published benchmark: 300 examples over AG News, Banking77 and Emotion, all Choice.

Letter-readout model. A decision model that answers with an option letter after its own prompt (JevK5, plumb-4b, Quyet, spark-s1), read from the letters' logits as its runtime reads them. More.

Noul. A yes/no question. Answers noul, the probability of yes.

Option. One possible answer of a question: a key and a description.

Pointer reading. How a decision model answers: a trained head over delimiter tokens of its own format.

Preset. A model plus its reading settings (layers, prompts, temperature, center), built in or fitted by jul models add. More.

Reading. How a model is turned into an answer: vector, cross, pointer, contrastive, letters, a tuned head. More.

Score. A question that places the text on an ordered scale. Answers score, the expected level.

State. The text (or object) a call asks about.

System One. Jev's name for its typed-decision API: POST /v1/systemone. jul serve speaks it, SystemOneHTTP calls it.

tau (τ). The temperature of softmax(cosine / tau), fitted per model on development sets.

Zero-shot. Answering with no example of the task: only the question and its options.