Datasets:

---

Qwen

/

BlenderWorld

Like 64

Follow

Qwen 113k

Hi @VoyagerXHF @zhibogogo @nebulae09 👋

I'm Nathan from Hugging Face. I lead "Leaderboards on the Hub". Congrats on releaseing BlenderWorld!

How it works:

1. Add an eval.yaml at the root of this repo. Since BlenderWorld already ships as a Harbor package, it can use evaluation_framework: harbor.
2. We allow-list it, and the dataset gets a "Benchmark" tag.
3. Results show up automatically: scores published in model repos (.eval_results/) appear as a leaderboard on this dataset page and on each model's page, linking back to BlenderWorld.

You stay in control of the benchmark and of how results appear. Docs: https://huggingface.co/docs/hub/eval-results · https://huggingface.co/docs/hub/leaderboard-data-guide

The results in your paper (Table 2) would be a great starting point, especially for the open-weight models (DeepSeek, Qwen3.8, GLM, Kimi, MiniMax).

Happy to open the eval.yaml PR for you so you only have to review and merge. Let me know if you'd prefer a different headline metric (overall Score vs. Text / MM).

Thanks!
Nathan