Embeddings bpefilesloader load

BpeFilesLoader.Load

Reads a vocab.json and merges.txt pair into a BPE vocabulary.

public static BpeVocabulary Load(Stream vocabJson, Stream merges, ArtifactLoadOptions options = null, bool byteLevel = true)
public static BpeVocabulary Load(string vocabJsonPath, string mergesPath, ArtifactLoadOptions options = null, bool byteLevel = true)

Parameters — vocabJson and merges are the two streams, never disposed here; vocabJsonPath and mergesPath are the two files. options bounds what will be accepted and defaults to ArtifactLoadOptions's own defaults. byteLevel says whether the model is a byte-level BPE, which GPT-2 and its descendants are.

Returns — BpeVocabulary, with the merge list in the order the file gave it.

Exceptions — ArgumentNullException for a null source or path. InvalidDataException when the content is not the format expected, declares a model this loader does not read, or exceeds a bound in options — the message names both the limit and the value.

Example — the GPT-2 layout, both halves from the same checkpoint.

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

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

Remarks — The two files must come from the same checkpoint. Nothing here can detect that they do not: a vocabulary from one model and merges from another load without complaint and tokenize wrongly, because the merge ranks address pieces that mean something else.

Merge order is the algorithm — the lowest-ranked applicable merge is applied first — so the file's line order is significant and a sorted merges.txt is a different model.

byteLevel stays a parameter because neither file records it. That is the standing argument for preferring TokenizerJsonLoader.LoadBpe when the checkpoint ships a tokenizer.json: there, the flag is read from the file.

Applies to — net10.0, netstandard2.0.

See also — BpeFilesLoader, BpeFilesLoader.LoadAsync, the persistence index.