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.

BpeFilesLoader

Reads the vocab.json + merges.txt pair GPT-2 ships, the layout that predates tokenizer.json.

public static class BpeFilesLoader

Example — two files, one vocabulary.

using Lodestar.Embeddings.Persistence;
using Lodestar.Embeddings.Tokenization;

BpeVocabulary vocab = BpeFilesLoader.Load("gpt2/vocab.json", "gpt2/merges.txt");
var tokenizer = new BpeTokenizer(vocab);

Remarks — two files because BPE is two things: vocab.json maps a piece to an id, and merges.txt gives the ordered merge list. Order is the algorithm — the lowest-ranked merge applies first — so the pair has to come from the same checkpoint. Mixing a vocabulary from one model with merges from another loads without complaint and tokenizes wrongly.

byteLevel defaults to true, which is GPT-2 and its descendants. Load explains when it is not.

Neither file records whether the model is byte-level, which is why that stays a parameter here and is read from the file by TokenizerJsonLoader.LoadBpe. Prefer the tokenizer.json route when the checkpoint offers one.

Reference behaviour is tokenizers.models.BPE.from_file(vocab, merges).

Applies to — net10.0, netstandard2.0.

See also — TokenizerJsonLoader, ArtifactLoadOptions, the persistence index.

Members

Member	What it does
BpeFilesLoader.Load	Reads a vocab.json and merges.txt pair.
BpeFilesLoader.LoadAsync	The same, asynchronously.