Datasets:
Add eval.yaml to register as a Hub benchmark
Merged, thank you Nathan. Two questions so the benchmark goes live: is there anything else we need to do for the allow-list? And should we add an evaluation_framework value to eval.yaml? The spec lists it, the PR doesn't set one, and our scoring is custom (accuracy, KL, Brier and ECE over full probability distributions, one request per case), so we are not sure which identifier fits. Our own hosted entries have no model repo, so we can point the open-weight entries at .eval_results once the page is live.
Hey! Just added the bench to the allow list. And added the results files for the models that were on the hub.
Would be great to benchmark on some more models available on the hub like laya!
This is a really interesting time for decision models as everyone is trying to crack the recipe, this benchmark will be super useful to guide the development of those models.
Here is a list of models, most of them on the hub already: https://multimodalart-jev-decision-index.static.hf.space/index.html
Thanks, Nathan. I opened a call for results (#10) so the authors can add their own scores, and I'll point the main ones to it. One question: for models you've already added results for, should authors leave them as they are, or would you prefer they open the PRs themselves? I also opened PRs for Jeff-Qwen3.5-0.8B and Jeff-Qwen3.5-2B before I saw your files, so please close mine if they duplicate yours.
Two quick questions about the leaderboard: does it need a refresh or just time to index, and are fine-tuned models hidden by default until a filter is switched? Our MiniLM specialist (adaptive-classifier/typed-decisions-minilm-l6-specialist) is a fine-tune, and its .eval_results are on main and read by the Hub API, but the leaderboard still shows no rows for any model.