witcheer/hermes-pairing-bench
收藏资源简介:
这是一个名为Hermes Pairing — Agentic Benchmark for Local LLMs (Phase A)的数据集,用于评估本地大型语言模型驱动代理的能力。数据集通过将本地模型与Hermes Agent(一个CodeAct代理)配对进行测试,该代理使用Python代码编写来协调工具,而非生成JSON函数调用。数据集包含一个排行榜,基于四个轴(加权总和为Hermes Pairing Score,范围0-100)对模型进行排名:codeact(代码作为动作,权重0.40)、longcontext(长上下文检索,权重0.25)、instruction(指令遵循,权重0.20)和multistep(多步循环稳定性,权重0.15)。数据在NVIDIA RTX 5090(32GB)上使用llama.cpp/GGUF和真实约3.5K令牌系统提示生成,属于可重复的合成排名阶段(Phase A),用于评估代理能力,而非实际验证(Phase B为验证步骤)。数据集还包括模型性能分析、方法细节和关键发现,如模型效率对比、长上下文内存限制等。
This dataset, named Hermes Pairing — Agentic Benchmark for Local LLMs (Phase A), is designed to evaluate how well local large language models can drive an agent. It pairs local models with Hermes Agent (a CodeAct agent by NousResearch), where the model acts by writing Python code (`execute_code`) to orchestrate tools, rather than emitting JSON function calls. The dataset includes a leaderboard ranking models based on four axes (with a weighted sum called Hermes Pairing Score, ranging 0–100): codeact (code-as-action, weight 0.40), longcontext (retrieve-and-use in long context, weight 0.25), instruction (compliance under real prompt, weight 0.20), and multistep (loop stability, weight 0.15). Data was generated using llm-bench-rig on an NVIDIA RTX 5090 (32GB) with llama.cpp/GGUF and under Hermess real ~3.5K-token system prompt. It represents a reproducible synthetic ranking phase (Phase A) for agentic capability assessment, with Phase B serving as the validation step using the actual Hermes Agent. The dataset also provides findings, methodology details, and model information, highlighting aspects like model efficiency, long-context limitations, and performance trade-offs.




