Skip to content

Request to release the agent-SFT target checkpoint for reproducing “Agent SFT + Harness-R1” #2

Description

@wxxv

Dear author,
Thanks for releasing Harness-R1—the code, patch protocol, and both engineer checkpoints have been very helpful.

I am trying to reproduce the Agent SFT + Harness-R1 row in Table 1 (59.2 → 64.2). The README notes that agent-sft-harness-r1 is only meaningful against the frozen agent-SFT Qwen3.5-9B target it was trained on, but that target is not in the release. Without the same weights, we cannot pair the public engineer with a matching actor, so the number is not reproducible even though the engineer itself is available.

Would you consider releasing the agent-SFT target checkpoint (the Qwen3.5-9B after the 2,515-trajectory SFT in Appendix B.3)?

Thank you for considering this.

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

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions