rostlabs/rost-286m-diacritics-stripped

286M parameter nanochat GPT (base stage), research tag stripped-d12, checkpoint step 001920.

Part of the rost model zoo — the full set of trained variants behind the rost research series, published for reproducibility. This is a research model, not a product.

architecture	nanochat GPT, 12 layers, 768 embed, 2048 context
tokenizer	included under tokenizer/ (32,768 vocab)
training mixture	Romanian with all diacritics stripped
stage	base
val bpb (own split)	0.84298

Diacritics-study ablation (research tag stripped-d12): the stripped twin of rost-286m-base. The diacritic penalty grows with scale — ~60% at this size — which is why the rost recipe keeps the diacritics.

Validation bits-per-byte is measured on this arm's own validation split and is not comparable across arms — cross-arm comparisons in the rost write-ups are always cross-evaluated on identical text.

Licence

CC-BY-NC-4.0, non-commercial, inherited from the most restrictive component of the training data.