rich6/tenacious-bench-v0.1
收藏资源简介:
Tenacious-Bench v0.1是一个包含200个任务的机器可验证基准数据集,专为B2B销售邮件代理设计,用于评估其在工程招聘领域的特定政策检查和语调标记。数据集分为训练集(97个任务)、开发集(60个任务)和保留集(43个任务),覆盖了7种失败维度和4种创作模式。每个任务代表一个B2B销售场景,包含输入、候选输出、真实标签和评分标准。数据集通过程序化生成、跟踪衍生、多LLM合成和手工编写四种方法创建,确保多样性和真实性。所有数据均为合成或经过匿名处理,不包含真实个人信息。
Tenacious-Bench v0.1 is a 200-task machine-verifiable benchmark for B2B sales email agents, graded on 6 deterministic policy checks and 5 LLM-scored tone markers derived from the Tenacious Consulting style guide. The dataset is partitioned into 97 train, 60 dev, and 43 held_out tasks, covering 7 failure dimensions and 4 authoring modes. Each task represents a B2B sales outreach scenario with input, candidate output, ground truth label, and scoring rubric. The dataset was created using four methods: programmatic generation, trace-derived adaptation, multi-LLM synthesis, and hand-authored adversarial examples. All data is synthetic or redacted, with no real prospect information included.





