Datasheet
Motivation
RoutingBench V2-beta isolates dynamic next-agent routing from actor-backbone quality. It evaluates whether a router uses the task, observable workflow state, execution feedback, budget, and qualitative role contracts to choose the next agent.
Composition
- 120 interactive deterministic workflows.
- 809 isolated decision states derived from reachable trajectories.
- 24 qualitative agent profiles.
- Eight workflow families and five pool-perturbation modes.
The release contains three Hugging Face configurations: decision,
interactive, and agent_catalog.
Collection and generation
The dataset is programmatically generated with seed 42 from original workflow templates. Existing multi-step benchmarks and project traces informed abstract failure patterns only. No original questions, answers, evidence passages, or attachments from those sources were copied.
Splits
Scenarios are split by split_group_id. States derived from the same workflow
template group cannot cross development, validation, and test splits.
The V2-beta test split is public and diagnostic. It is not a hidden test set.
Personal and sensitive information
The tasks are synthetic and are not intended to contain personal or sensitive information.
Quality control
- Deterministic regeneration and SHA-256 checksums.
- JSON schema and invariant validation.
- Zero observed-state label collisions.
- Oracle Track A admissible accuracy and Track B completion of 100%.
- Quotas for singleton, set-valued, trade-off, terminal, and failure states.
Maintenance
V2-beta is a public research preview. Incompatible changes require a new version; released artifacts should remain immutable and be referenced by Hub revision/tag.