Datasets:
Interactive explorer for the leaderboard: ranking + Pareto charts, reads this card live
Hi! First of all, congratulations on Typed Decisions. I really like this benchmark: soft gold distributions, calibration metrics next to accuracy and the teacher self-agreement reference make it far more informative than a plain accuracy table, and the way you keep zero-shot and fitted rows apart is great.
I built a small static Space to browse the leaderboard: https://huggingface.co/spaces/tardellirs/typed-decisions-explorer
- It reads this README on every visit and parses both tables in the browser, so new rows show up as soon as you merge them. Nothing to keep in sync on your side.
- Two Pareto charts: accuracy × size and accuracy × p50 latency, with switchable axes (KL, Brier, ECE, price). Frontiers are computed separately for zero-shot and fitted rows, and for API round trips vs local runs on the latency chart.
- The card's reference lines (prior, perfect factor recovery, teacher self-agreement) and Jev's score as a baseline.
- Sortable table and CSV export. Hovering a model shows an approximate 95% interval on accuracy, and rows within noise of the top one are marked.
If you find it useful, feel free to link it from the card. Feedback is very welcome.
Thanks for maintaining this!
Thanks for building this. Reading the README on each visit means it stays current with nothing to sync on our side, and keeping separate frontiers for zero-shot and fitted rows, and for API versus local latency, is the right call. The noise markers on rows near the top are a nice touch too.
Thanks, Sandeep! Glad it's useful. If the card's table format changes at some point, feel free to ping me here and I'll update the parser. Congrats again on the benchmark.