Model Router

August 26, 2026 · View on GitHub

Concrete data backing the Model Router PR to RouteWorks/RouterArena. For the conceptual background (TLD taxonomy, cal/test split, Pareto-optimal policy selection), see model-router.md.

TLD Policy Used in This PR

../data/routing_policy_for_router_arena.yaml — the {Task × Domain × Level → LLM} routing policy, using the mmBERT TLD classifier.
../data/cal/router_data.json — the 1,976-query calibration set used to fit this policy.
routing-heatmap.html — interactive heatmap of this policy's TLD → model mapping, calibration vs. test cell counts (open in a browser).

Pareto-Optimal Result

TBC — quality/cost numbers for the selected Pareto-optimal point (cal vs. test), compared against the best single model, will be added here.

TLD Classifier Weights

TBC — the mmBERT-based TLD classifier weights will be published on Hugging Face; the link will be added here once uploaded.