Skip to content

Support mixed-source drafting when one source shares a language code with the target #1088

Description

@bhartmoore

When at least one source text shares a language ISO code with the target text, mixed-source drafting fails. This ClearML experiment is one example.

This experiment seeks to combine a majority-language Bible with an English back translation on the source side, and translate into a language that shares a language code with the source Bible. It also uses include_test: false for the English back translation, which may be relevant to the failure. This limits testing to the source and target that share a language code. It's important to be able to exclude the back translation from testing in order to get meaningful model scores, because drafting from the back translation does not reflect the real-world task.

Notes: This issue does not occur frequently and is low-priority. Ideally any fix would also support many-to-many mapping where at least one source project shares an ISO code with the target.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    enhancementNew feature or request

    Projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions