Home/Modules/large-language-models

large-language-models

library module lib/large-language-models

lib/large-language-models

SLeeLa standard-library mirror of the artificial-intelligence/large-language-models C/C++/Python subsystem. These classes expose the same contracts in SLeeLa; the native sources remain authoritative for the actual kernels, schemas, and weights handling. Nothing here ships a model or weights.

SLeeLa classMirrors (native prefix)Purpose
SLPackage—Package facade for the compiler, Loader, and Nordshrift.
SLLlmEnginesllm_Thin engine over a caller-supplied backend adapter: load / unload / generate lifecycle and status codes.
SLLlmGenerationOptionssllm_generation_optionsDecoding knobs: max new tokens, temperature, top-p, seed.
SLLlmVocabularysllmv_Token-id store: stable insertion-order ids, add / find / token lookup.
SLLlmBaseVocabularysllmbv_Seed vocabulary loader (sleela-base-en-1.0): ten specials at fixed low ids, fail-closed id verification.
SLLlmReasoningsllmr_Reasoning Language Model (RLM): subjects, step kinds, verdicts, and the UNKNOWN-never-promoted commit policy.
SLLlmCapacitysllmc_Model-capacity profile (2–16 GB envelope) and derived reasoning budget per tier.
SLLlmCodeModelsllmcx_Procedural code-model profile: languages, task flags, verification check mode.
SLLlmSleelaVerifiersllmsl_Static SLeeLa-rule checker (reserved words, close() collision, comment forms, 64-field cap, metadata) that acts as the reasoning layer's symbolic half.
SLLlmParametrysllmp_Model binding: known-by-name (registry) or local-by-path, fail-closed.
SLLlmLoadersllme_Loader/handler idiom: fixed-width encode with OK / SKIP / ERROR, careful by construction.

A runnable demo lives in examples/LlmReasoningDemo.sleela.

Every class carries the standard Metadata Class record (see .kiro/steering/metadata-class-standard.md and SLEELA.syntax §15).

Files 13