Neural Studio 008

Language base: SmolLM2-360M-Instruct, Hugging Face / HuggingFaceTB.
https://huggingface.co/HuggingFaceTB/SmolLM2-360M-Instruct
Revision a10cc1512eabd3dde888204e902eca88bddb4951. Apache License 2.0.
The hosted editor graph adds trained LoRA attention branches, last-token
projection and a greedy token readout. The frozen pretrained tensor bytes
are downloaded from the upstream model, unchanged. Training data consists
of locally authored procedural scene programs; no images or videos were
used for this adaptation. Model training was not performed from scratch.

Runtime libraries: ONNX Runtime (Microsoft, MIT), Transformers.js (Hugging
Face, Apache 2.0). Their bundled license notices are in llm-worker.js.
Source, dataset generator, training/export/evaluation code and neural engine
are available in the source archive linked from the studio.