Lodestar.Embeddings 0.6.0. 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.

WordPieceVocabulary

The vocabulary and the settings that decide how it is read.

public sealed record WordPieceVocabulary

Properties — Vocab maps token to id. UnkToken is the token a word that cannot be covered becomes. ContinuationPrefix marks a piece that continues a word, ## in BERT. Lowercase says whether text is folded before matching. AddedTokens are the literal matches applied first. Count is how many entries the vocabulary holds.

Example — loading one from the vocab.txt a model ships.

using System.Text;
using Lodestar.Embeddings.Persistence;
using Lodestar.Embeddings.Tokenization;

var bounds = new ArtifactLoadOptions();
byte[] file = Encoding.UTF8.GetBytes("[UNK]\ntoken\n##ize\ntext");

WordPieceVocabulary vocabulary = VocabTxtLoader.Load(
    new MemoryStream(file), bounds, unkToken: "[UNK]", continuationPrefix: "##", lowercase: true);

int count = vocabulary.Count;  // => 4

Remarks — a vocab.txt is one token per line and the id is the line number, which is why loading it needs no ids and why editing such a file by inserting a line renumbers everything after it. That is a real way to break a model quietly.

The settings travel with the vocabulary rather than with the tokenizer because they are properties of the file: a vocabulary trained lowercase cannot be read case-sensitively, and pairing it with the wrong ContinuationPrefix produces tokens that exist nowhere in it.

Applies to — net10.0, netstandard2.0.

See also — WordPieceTokenizer, AddedToken.

Members

Member	What it does
WordPieceVocabulary.Equals	Value equality over the entries and settings.
WordPieceVocabulary.GetHashCode	A hash consistent with it.