DecisionBench Leaderboard

The public leaderboard and result explorer for DecisionBench, the benchmark suite for general-purpose decision models.

This project is a maintained fork of the Apache-2.0 licensed MTEB leaderboard frontend. It preserves MTEB's benchmark, task, model, comparison, filtering, charting, accessibility, and deep-linking facilities while adapting the data contract and terminology to typed decision models.

Report leaderboard bugs and request features in the central DecisionBench issue tracker.

Development

npm ci
npm run dev

The reviewed result matrix is loaded from static/leaderboard.json. It is generated from Hanno-Labs/decision-bench-results, not edited by hand.

The results repository supplies each reviewed model_type. When a recipe's result URL points to different Hugging Face weights, the UI keeps the recipe's credited name and shows those weights separately. Older data snapshots without a reviewed type use the legacy type heuristic.

Verification

npm run check
npm run lint && npm run test:unit && npm run build

Deployment

The included Dockerfile builds a static SvelteKit application and serves it on port 7860, suitable for Hugging Face Spaces. GitHub Actions also verifies the application and publishes the container image.

Upstream and license

DecisionBench Leaderboard tracks embeddings-benchmark/leaderboard-frontend as its upstream. The original Apache-2.0 license is preserved in LICENSE.