Machine-verifiable evaluation benchmark for B2B sales-agent quality. 257 tasks · four authoring modes · three sealed partitions · SimPO-trained judge adapter.

Dataset Summary

Tenacious-Bench measures whether a (brief, email) pair satisfies five rubric dimensions that are fully machine-verifiable without an LLM:

Code Dimension Rule

D1 grounding_fidelity All numerics in email traceable to brief

D2 icp_pitch_alignment Ambiguous segment → no… See the full description on the dataset page: https://huggingface.co/datasets/Birkity/tenacious_bench_v0.1.