Skip to content

data ingest: remove the tokenizer.json ingest requirement for MLM (RFC-0002) #184

Description

@LukasWodka

RFC-0002 §12.11. Remove the tokenizer.json ingest requirement for masked_language_modeling.

Verified: the ingestor never reads it, the schema never required it, and the training side deliberately removed the dataset-staged tokenizer in #805 (it caused vocab divergence that broke weight averaging). The tokenizer is the collaborator's, supplied at model upload.

  • Delete the MLM tokenizer branch in DiscoverText (stop requiring/staging it), drop the manifest tokenizer field, fix the false "the ingestor reads it" comment, update the tests. CLI-only; low risk.

Design-of-record: RFC-0002 (#174). Part of the data-ingest redesign epic.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Fields

    No fields configured for issues without a type.

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions