Teach small models new tricks.

01 / Portable helpers, directly in Noqeri

Lib/ai/framework.noqeri exports scalar linear prediction, squared-error gradients, bounded SGD steps, pair counting, attention score and guarded agent budgets. This source can be checked by the normal compiler without auto-linking the native tensor SDK.

import "../../Lib/ai/framework.noqeri"
let loss: float = ai_framework_loss(0.0, 0.0, 2.0, 4.0)
let weight: float = ai_framework_sgd_weight(0.0, 0.0, 2.0, 4.0, 0.1)
print(loss)
print(weight)

02 / Native tensor and autodiff SDK

The optional libnoqeri_ai C++17 library includes bounded dense tensor operations, automatic differentiation, CPU scalar/tiled/SSE2 backends and explicitly bound Noqeri AOT entry points. Importing a .noqeri file does not automatically link this library.

03 / Experimental local transformer trainer

Tools/ai and Runtime/native include a bounded causal transformer research trainer with seeded tokenization, checkpointing, SGD/AdamW, limited stacked layers and selected fp16/fp32/fp64 compute paths. It is a local preview, not a scalable foundation-model training platform.

noqeri ai transformer train corpus.txt scratch.nqtf --steps 200 --context 4 --hidden 8 --merges 8 --backend auto
noqeri ai transformer generate scratch.nqtf "hello" --tokens 24

04 / GPUs require actual hardware proof

An optional OpenCL acceleration path and FP precision variants exist, but physical-device availability, backend accuracy and benchmark speedups are separate tests. Enabling a CMake flag or compiling host code does not prove that a GPU kernel executed.

05 / Safe agents and model boundaries

Lib/ai/agent.noqeri provides bounded policy and step helpers. Untrusted shell, arbitrary network and external tool access are not silently granted; embedding hosts must explicitly authorize capabilities. Full Unicode tokenization, production distributed training and general GPU coverage are open.