What is Ollaya, really?

[eightkpage]

How do the decision models differ from LLMs?

[lunar-me]

Ollaya runs open decision models locally, the way Ollama runs LLMs. A decision model reads a STATE (a message, an email, a ticket, generally any JSON structure) plus typed questions ('choice' or 'score') and returns calibrated probabilities in a single forward pass, in milliseconds, it never generates any text but generates CHOICE.

[cobanov]

Adding to @lunar-me's answer, which is right in substance:

• There are three question types: choice (one of 2–255 labels), score (a scale of 2–10 levels; the answer is the expected level) and noul (the probability that a statement holds).
• The model doesn't generate a choice. It gives each option you listed a probability, calibrated with the model's fitted temperature, and choice is the most likely label.

On how they differ from LLMs: many of them are LLMs. winnow, kev, decider, jeb and clef are Gemma and Qwen language models fine-tuned for this, and Ollaya reads them differently: one forward pass over a fixed prompt, then the score the model gives each option's label (or a small head on its hidden states), instead of sampling text. Others aren't LLMs at all: laya, von, decima and nli are BERT-style encoders that read the whole request at once and score every option. Either way you get a fixed answer shape with no parsing and no label you didn't offer, probabilities you can threshold (act above one value, escalate below), and speed: 7–35 ms for five questions on a GPU for the encoders, 0.1–0.9 s for most decoders. What you give up is any explanation and any answer outside your options; for those, use an LLM, or escalate the low-confidence cases to one.

More in the FAQ.