Submission review request: Qwen2.5-Omni-7B with-obs SFT (Pass@1 = 15.76)

#2
by ZhangYuchi - opened

Hi maintainers,

I submitted my Qwen2.5-Omni-7B fine-tune 8 days ago (commit 42995fb, path
submissions/2026/06/Qwen2_5-Omni-7B-with-obs-SFT__ms-swift__20260630_094618.json).
The submission has been auto-verified but is not yet published on the leaderboard
(results.json last updated 3 months ago when Orchestra-o1-GPT-5 was added).

Summary of the submission:

  • Method: Qwen2.5-Omni-7B + with_obs SFT (ms-swift)
  • Public model weights: https://huggingface.co/ZhangYuchi/modelbest-Qwen-2.5-Omni-7B-SFT-only
  • Self-reported Pass@1 (4 seeds): 42→16.94, 1337→15.28, 2024→14.44, 7→16.39
    mean 15.76 ± 1.12
  • Eval protocol: official run_base_agent_original_paper.py with
    DeepSeek-V3.2 as LLM judge, --enable-active-perception, all default tools.
  • Training objective: functionally equivalent to the paper's
    trajectory-level masked SFT (only assistant tokens contribute to the loss;
    observation tokens are masked via role normalization + --loss_scale default).
    See the model card for full details.
  • Contamination self-check: 0/2151 verbatim question overlap between
    train and test.

Just wanted to give a friendly ping in case my submission slipped through.
Happy to help with re-verification, run additional seeds, or answer any
question — please let me know.

Thanks!

NLPIR Lab @ RUC org

Thanks for your interest in OmniGAIA! We're glad to see you achieved positive performance gains.

We've reviewed your submission. The entries "method_name":"Qwen2.5-Omni-7B-with-obs-SFT (ms-swift)" and "organization":"modelbest" are a bit unclear to us. Could you please provide the official method name and your actual institutional affiliation?

Hi, thanks for getting back to me and for reviewing the submission!

Sorry for the ambiguity in the metadata. Let me clarify both fields:

  • method_name: Please use "MB-Omni-SFT". The core contribution is a trajectory-level masked SFT recipe on top of Qwen2.5-Omni-7B, where observation tokens are kept in context but excluded from the loss (via role normalization + --loss_scale default in ms-swift). The "(ms-swift)" suffix was just the training framework and can be dropped from the display name.

  • organization: Please use "ModelBest Inc. (面壁智能)". I'm affiliated with ModelBest as an intern, and this submission was produced using ModelBest's training resources, so attributing it to ModelBest is appropriate.

Everything else (model weights link, eval protocol, seeds, contamination check) remains as reported in the original post. Happy to re-run additional seeds or provide any further information if that helps verification.

Thanks again!

Small update on top of the SFT clarification above — I've also submitted a follow-up DPO checkpoint just now:

Full 4-seed results for transparency (same eval protocol as the SFT submission):

Seed SFT Pass@1 DPO Pass@1
7 16.39 15.00
42 16.94 18.06
1337 15.28 15.28
2024 14.44 17.50
Mean ± std 15.76 ± 1.12 16.46 ± 1.55

The DPO improvement is not uniform across seeds — it mainly helps weak seeds (e.g. seed 2024: +3.06) while preserving the strongest seed (seed 42: +1.12). Happy to provide the other 3 seed JSONs or any additional info if needed.

One small housekeeping note: I accidentally drag-and-dropped the DPO file twice, so submissions/2026/07/ currently has two identical copies:

  • MB-Omni-SFT-DPO_20260722_022717.json (duplicate, please remove)
  • MB-Omni-SFT-DPO_20260722_022724.json (keep this one)

I don't have delete permission on the repo, so could you please remove the earlier duplicate when you get a chance? Sorry for the noise!

Thanks!

NLPIR Lab @ RUC org

Thanks for your submission. We’ve merged your results.

Thank you!

Sign up or log in to comment