provost-model
Copyright 2026 Martin J. Gallagher

This product includes a derivative of SmolLM2-135M
(https://huggingface.co/HuggingFaceTB/SmolLM2-135M), Copyright Hugging Face,
licensed under the Apache License, Version 2.0.

Changes made to SmolLM2-135M:
- fine-tuned with a LoRA adapter (rank 16, every linear layer) and two
  classification heads on line-layout labels of command output, using
  provost's ml/train_lm.py;
- the adapter merged into the base weights;
- cut to the first 18 of its 30 transformer layers;
- the language-modelling head left out (the heads above replace it);
- stored at 8 bits per weight as numpy arrays, by provost's
  ml/export_lm.py.

tokenizer.json is SmolLM2-135M's tokenizer, as the transformers library loads
it, unchanged.
