FlowBench / harbor /README.md
jwu323's picture
Clarify FlowBench Harbor scoring boundary
bd5cad6 verified
|
Raw
History Blame Contribute Delete
1.44 kB

FlowBench Public Harbor Pack

This is the answer-free public Harbor task package. It is suitable for inspecting tasks, integrating agents, and launching Harbor containers with a smoke verifier. It is not the official scoring package.

Scope note: this package is a business-operations tool-composition benchmark. It is unrelated to prior workflow-guided planning or workflow-generation benchmarks that also use the FlowBench name.

Files intentionally excluded from this public package:

  • fixed gold labels
  • Harbor oracle solutions
  • strict verifier expected files
  • model result logs and generated transcripts

Official-style or private benchmark scores should be computed with a private evaluator pack or a freshly generated held-out split. Publishing fixed labels for the official split would make hard-coded benchmark submissions possible.

For public runs, freeze prompts and adapters before scoring and reject task-id-specific lookup logic. Public Harbor results are smoke-test results, not official FlowBench scores.

Data source:

  • There is no separate CSV, database dump, or network service. Each Harbor task includes /app/flowbench_tools.py, which deterministically builds the same synthetic customers, products, orders, returns, tickets, inventory, FX rates, and SLA policies on every machine.

Harbor smoke run, replacing <agent> with the Harbor agent to evaluate:

harbor run -p harbor -a <agent> -l 1