Skip to content

fix(grpo): coordinate precision loss in fraction-pixel roundtrip #49

Description

@abrichr

Problem

_parse_vlm_output_to_action converts fractions to pixels with int() truncation (trainer.py:767-768), then _format_action_as_text converts back with 2-decimal formatting (trainer.py:707). External RL training use case models use 3 decimal places (e.g., 0.461). The roundtrip loses precision:

  • 0.461 * 1920 = 885.12 -> int(885)
  • 885 / 1920 = 0.4609375 -> "0.46"

Proposed Fix

Consider storing coordinates as fractions throughout and only converting to pixels at the adapter boundary.

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

    bugSomething isn't working

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions