tau2 · README.md source

tau

Results of cap-evolve runs on tau2: the τ²-bench customer-service agent benchmark (airline / retail domains) and its cap-evolve variants. One of the per-benchmark result repos in the skillberry-runs org.

On the dashboard: https://skillberry-runs.github.io/dashboard/tau/. The run rows (reward, cost, models per run) are on the main table (filter Benchmark); each row's Open UI opens that run's CapEvolve dashboard from this repo's ci_ui/.

What is here

bench value in the run records runs
tau2 47
tau2_custom_blackbox 6
tau2_custom_direct 3
tau2_custom_spa 3
tau2_sierra 1
tau2_sierra_blackbox 2
tau2_blackbox 1
skillberry_tau2_direct 2
skillberry_tau2_spa 2

Adding to it

See the dashboard's PUBLISHING.md: section B for write-ups, F for "Open UI" data.