Skip to main content

Crate cerno_core

Crate cerno_core 

Source
Expand description

The cerno engine: three typed primitives on top of one token of model output.

A Noul, a Choice and a Score are the same operation wearing different names. Each offers the model a handful of lettered options and reads the probability distribution over the first token it would generate. That is one forward pass, no matter how many options there are — which is what makes the whole thing fast, and what separates it from asking a model to write an answer and then parsing the prose back out.

  question ──> options ──> labels A,B,C ──> prompt ──> host ──> distribution
                                                                      │
           typed answer <── calibrate <── fold variants, floor <──────┘

The pieces are split so each can be tested on its own: labels owns the alphabet, prompt owns the wording, math owns the arithmetic, and engine joins them.

Re-exports§

pub use engine::Engine;
pub use engine::EngineError;

Modules§

engine
The one path all three primitives take.
labels
The answer alphabet.
math
Turning label logprobs into probabilities and a confidence.
prompt
Prompt construction.