Embeddings 0.4.0 precompilednormalizer

Lodestar.Embeddings 0.4.0. This page is frozen at that release. Read the current documentation for what main says now. A link to a decision or a migration page follows main, and leaves the archive.

PrecompiledNormalizer

SentencePiece's charsmap normalization: the character folding a model was trained with.

public sealed class PrecompiledNormalizer

Properties — CharsMapLength is the size in bytes of the compiled trie it was built from.

Example — the normalizer a model shipped, applied before tokenizing.

using Lodestar.Embeddings.Tokenization;

// charsMap is the precompiled_charsmap blob from the model's own spiece.model.
byte[] charsMap = File.ReadAllBytes("spiece.model");

PrecompiledNormalizer normalizer = PrecompiledNormalizer.FromCharsMap(charsMap);

string folded = normalizer.Normalize("Ｈｅｌｌｏ");

Remarks — this is not Unicode normalization by a named form. It is a trie compiled into the model file, mapping character sequences to replacements, and it encodes decisions the model's authors made — full-width to half-width, some accents stripped, some kept. Substituting NFKC for it gives text the model was not trained on.

A vocabulary's Normalizer is null when the model ships no charsmap, and then the text reaches the tokenizer unchanged. A stock T5, ALBERT, XLM-R or camemBERT ships one; a model built by hand usually does not.

Every fence on these pages is docs-run: skip for the same reason the ONNX pages are: the input is a binary artifact and none is committed. They are still compiled, so a renamed member still fails CI.

Applies to — net10.0, netstandard2.0.

See also — SentencePieceVocabulary, SentencePieceTokenizer.

Members

Member	What it does
PrecompiledNormalizer.Equals	Value equality over the charsmap.
PrecompiledNormalizer.FromCharsMap	Build one from a model's compiled trie.
PrecompiledNormalizer.GetHashCode	A hash consistent with it.
PrecompiledNormalizer.Normalize	Apply the folding to one string.