Skip to content

Running zero shot gemini flash for tibetan text layer detection task. #421

Description

@karmatai

Description

In this card we are going run zero shot gemini flash to detect tibetan text layers(Quotation, tsawa and yigchung) and see whether llm beats using encoder in this current task.

Subtask

  • Built window on the test data based on gemini tibetan token size and span length.
  • Run smoke text on gemini flash 3.7 and 3.8 to check quality per dollar.
  • Perpare prompt for both task( Quotation ,tsawa) and yigchung to get the exact result in text.
  • Recover offset from the span.
  • Price the full run and implement the run.
  • Evaluate the result.
  • Document the result.

Reviewer

Activity

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

Metadata

Metadata

Assignees

Labels

No labels
No labels

Type

No type

Fields

Priority

None yet

Projects

  • Status
    In Progress

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions