gdpval-evals
收藏资源简介:
GDPVal Hermes Agent vLLM 评估追踪数据集是一个私有的评估产物存储库,用于记录和分析Hermes智能体在本地通过vLLM服务在NVFP4和FP8量化大型语言模型上执行GDPVal(OpenAI)基准测试的结果。数据集包含每次评估运行的详细产出文件,如运行摘要、按任务统计的使用情况文件以及每个请求的遥测日志(存储于`agent_requests.jsonl`并归档于`telemetry.tar.gz`)。评估围绕220个任务展开,核心评估指标为平均奖励(`avg_reward`)。数据集提供了全面的模型性能矩阵,包括已完成评估的头条结果(涵盖Qwen、Gemma、Nemotron等多个模型系列的不同量化版本)以及因基础设施问题而处于待处理/部分完成状态的运行记录。此外,数据集详尽记录了每个模型在评估时使用的精确vLLM服务配置(Docker命令和参数),确保了评估环境的可复现性。除了基础性能分数,数据集还包含深入的性能分析指标:多令牌预测(MTP)草稿令牌的逐位置接受率分析、部分运行的llama.cpp引擎计时详情(如TTFT、吞吐量)、每次运行的元数据(如使用的代理、引擎、评判模型、硬件)以及详细的代理端令牌消耗统计。该数据集适用于研究大型语言模型在智能体任务中的性能、比较不同量化技术(如FP8、NVFP4、AWQ、GPTQ)的效果、分析推测解码策略的效率,以及评估服务配置对模型推理性能和资源消耗的影响。
The GDPVal Hermes Agent vLLM Evaluation Tracking Dataset is a proprietary evaluation artifact repository designed to record and analyze the results of GDPVal (OpenAI) benchmark tests executed by Hermes Agent on large language models quantized with NVFP4 and FP8 formats via local vLLM serving services. The dataset contains detailed output files from each evaluation run, including run summaries, task-wise usage statistics files, and telemetry logs for every request, which are stored in "agent_requests.jsonl" and archived in "telemetry.tar.gz". The evaluation covers 220 tasks, with the core evaluation metric being average reward ("avg_reward"). The dataset provides a comprehensive model performance matrix, including headline results of completed evaluations covering various quantized versions of multiple model families such as Qwen, Gemma, and Nemotron, as well as run records that are pending or partially completed due to infrastructure issues. In addition, the dataset thoroughly documents the exact vLLM service configurations (Docker commands and parameters) used for each model during evaluation, ensuring the reproducibility of the evaluation environment. Beyond basic performance scores, the dataset also includes in-depth performance analysis metrics: per-position acceptance rate analysis of draft tokens for Multi-Token Prediction (MTP), timing details of the llama.cpp engine for partial runs such as Time to First Token (TTFT) and throughput, metadata for each run (e.g., used agent, engine, judging model, hardware), and detailed token consumption statistics on the agent side. This dataset is applicable to research on the performance of large language models in agent tasks, comparing the effects of different quantization techniques such as FP8, NVFP4, AWQ, GPTQ, analyzing the efficiency of speculative decoding strategies, and evaluating the impact of service configurations on model inference performance and resource consumption.
数据集概述:GDPVal Hermes Agent vLLM Evaluation Traces
该数据集包含基于 OpenAI GDPVal 基准的 Hermes Agent 评估工件,所有评估均针对由 vLLM 服务的本地 NVFP4 和 FP8 模型进行。数据集本质上是一个工件存储库,而非经过清洗的行式数据集,其核心数据存在于运行摘要、按任务划分的使用文件以及按请求记录的遥测日志中。
核心指标
- avg_reward:所有 220 个任务的平均奖励值,是衡量模型性能的关键指标。
- pass_at_1:为与 tau2 参考模式保持向后兼容而提供,但在 GDPVal 上过于稀疏,无法有效区分模型(最佳运行约 5%)。建议使用
avg_reward及其分布。
数据结构
布局与 exolabs/tau2-hermes-nemotron-eval-data 一致:每个完成的运行位于顶层 <model-shortname>_<TS>/ 目录下。目录内包含运行摘要、按任务的使用文件、以及归档的每请求遥测日志(agent_requests.jsonl)。
主要评估结果(Headline Matrix)
下表展示了已完成评估且结果完整的模型性能排名。
| 排名 | 模型 | 任务数 | 平均奖励 | 中位数奖励 | pass@1 | 零奖励任务数 | 最大轮次命中 |
|---|---|---|---|---|---|---|---|
| 1 | Jackrong/Qwopus3.6-27B-v2 | 219 | 0.7203 | 0.7800 | 0.0320 | 4 | 6 |
| 2 | sakamakismile/Qwen3.6-27B-NVFP4 | 20 | 0.6447 | 0.8146 | 0.0500 | 2 | 0 |
| 3 | cyankiwi/Qwen3.6-35B-A3B-AWQ-4bit | 219 | 0.5873 | 0.7111 | 0.0228 | 43 | 4 |
| 4 | sakamakismile/Qwen3.6-27B-Text-NVFP4-MTP | 219 | 0.5837 | 0.7455 | 0.0183 | 39 | 1 |
| 5 | google/gemma-4-31B-it | 220 | 0.5663 | 0.5982 | 0.0091 | 10 | 0 |
| 6 | Qwen/Qwen3.6-35B-A3B-FP8 | 219 | 0.5533 | 0.6727 | 0.0365 | 43 | 8 |
| 7 | Qwen/Qwen3.6-27B | 219 | 0.5280 | 0.6667 | 0.0091 | 50 | 1 |
| 8 | leon-se/gemma-4-E2B-it-FP8-Dynamic | 219 | 0.1717 | 0.0000 | 0.0000 | 135 | 0 |
| 9 | leon-se/gemma-4-E2B-it-FP8-Dynamic | 220 | 0.1556 | 0.0000 | 0.0091 | 136 | 0 |
| 10 | nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4 | 219 | 0.0000 | 0.0000 | 0.0000 | 219 | 0 |
注:最大轮次命中 指因达到代理交互上限而截断的任务,这些任务被视为模型输出结果(奖励为0),而非基础设施故障。
待处理/部分运行结果
以下模型的评估因基础设施或实施故障而未能完全完成,目前被隔离。
| 模型 | 总任务数 | 已评分 | 待处理故障数 | 部分平均奖励 | 状态 |
|---|---|---|---|---|---|
| cyankiwi/gemma-4-26B-A4B-it-AWQ-4bit | 220 | 70 | 150 | 0.4732 | 存在基础设施故障 |
| cyankiwi/gemma-4-31B-it-AWQ-4bit | 220 | 220 | 0 | 0.5441 | 存在缺失任务 |
| leon-se/gemma-4-E4B-it-FP8-Dynamic | 219 | 219 | 0 | 0.1968 | 存在缺失任务 |
| openai/gpt-oss-120b | 219 | 218 | 1 | 0.4883 | 存在基础设施故障 |
| openai/gpt-oss-20b | 219 | 88 | 131 | 0.0000 | 存在基础设施故障 |
| LiquidAI/LFM2.5-8B-A1B | 219 | 219 | 0 | 0.0595 | 存在缺失任务 |
| mimo-v2-flash-iq2m | 20 | 11 | 9 | 0.0697 | 存在基础设施故障 |
| mimo-v2-flash-iq2m | 20 | 20 | 0 | 0.3227 | 存在缺失任务 |
| mimo-v2-flash-iq2xxs | 20 | 20 | 0 | 0.2563 | 存在缺失任务 |
| NVIDIA-Nemotron/Nemotron-Ultra-3-mopd-052726-mixed_nvfp4-fp8_amax_1000x65k | 220 | 71 | 149 | 0.6886 | 存在基础设施故障 |
| nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-FP8 | 220 | 220 | 0 | 0.3252 | 存在缺失任务 |
| nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-NVFP4 | 220 | 218 | 2 | 0.3345 | 存在基础设施故障 |
| nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4 | 220 | 7 | 213 | 0.0000 | 存在基础设施故障 |
| Qwen/Qwen3.5-122B-A10B-gptq-int4 | 220 | 5 | 215 | 0.0000 | 存在基础设施故障 |
| cyankiwi/Qwen3.5-9B-AWQ-4bit | 220 | 219 | 1 | 0.3289 | 存在基础设施故障 |
| lovedheart/Qwen3.5-9B-FP8 | 219 | 219 | 0 | 0.3999 | 存在缺失任务 |
| Qwen/Qwen3.5-9B | 220 | 6 | 214 | 0.2207 | 存在基础设施故障 |
| Qwen/Qwen3.6-27B-FP8 | 220 | 219 | 1 | 0.5830 | 存在基础设施故障 |
| sakamakismile/Qwen3.6-27B-NVFP4 | 220 | 219 | 1 | 0.6331 | 存在缺失任务 |
| RedHatAI/Qwen3.6-35B-A3B-NVFP4 | 219 | 217 | 2 | 0.4457 | 存在基础设施故障 |
| Qwen/Qwen3.6-35B-A3B | 220 | 217 | 3 | 0.6195 | 存在基础设施故障 |
| step-3.7-flash-iq4xs | 20 | 16 | 4 | 0.5853 | 存在基础设施故障 |
| step-3.7-flash-iq4xs | 20 | 20 | 0 | 0.4963 | 存在缺失任务 |
| step-3.7-flash-q3km | 20 | 20 | 0 | 0.6161 | 存在缺失任务 |
| step-3.7-flash-q3km | 20 | 20 | 0 | 0.5300 | 存在缺失任务 |
MTP 草稿令牌接受率
下表展示了采用 vLLM 引擎的模型在 MTP(Multi-Token Prediction)推理过程中的草稿令牌接受情况。
| 排名 | 运行标识 | 草案次数 | 平均接受长度 | p0 | p1 | p2 | p3 |
|---|---|---|---|---|---|---|---|
| 1 | nemotron-ultra-3-mopd-... | 4.22M | 4.28 | 0.925 | 0.858 | 0.786 | 0.711 |
| 2 | qwen3.6-35b-a3b_... | 3.88M | 3.84 | 0.836 | 0.742 | 0.665 | 0.601 |
| 3 | qwen3.6-35b-a3b-awq-4bit_... | 2.19M | 4.11 | 0.897 | 0.809 | 0.737 | 0.670 |
| 4 | qwen3.6-27b-text-nvfp4-mtp_... | 1.65M | 4.11 | 0.897 | 0.807 | 0.733 | 0.669 |
| 5 | qwen3.6-27b-fp8_... | 1.74M | 4.09 | 0.896 | 0.804 | 0.728 | 0.664 |
| 6 | qwen3.6-35b-a3b-fp8_... | 2.02M | 4.02 | 0.884 | 0.787 | 0.707 | 0.642 |
| 7 | qwen3.6-27b_... | 1.66M | 4.07 | 0.906 | 0.816 | 0.706 | 0.644 |
| 8 | gpt-oss-120b_... | 2.45M | 2.83 | 0.651 | 0.442 | 0.311 | 0.214 |
| 9 | qwen3.6-35b-a3b-nvfp4_... | 4.23M | 4.03 | 0.885 | 0.788 | 0.711 | 0.646 |
| 10 | qwen3.5-9b-fp8_... | 1.87M | 3.84 | 0.845 | 0.739 | 0.661 | 0.597 |
| 11 | qwen3.5-9b-awq-4bit_... | 2.14M | 3.68 | 0.823 | 0.702 | 0.613 | 0.540 |
| 12 | qwen3.5-9b_... | 1.14M | 3.64 | 0.815 | 0.691 | 0.603 | 0.533 |
| 13 | qwen3.5-122b-a10b-gptq-int4_... | 1.99M | 4.45 | 0.946 | 0.886 | 0.832 | 0.782 |
引擎时序(llama.cpp 按请求)
对于使用 llama.cpp 引擎的运行,记录了平均首次令牌时间(TTFT)、解码吞吐量、预填充吞吐量等时序数据。
| 排名 | 运行标识 | 平均TTFT (ms) | 解码吞吐量 (tok/s) | 预填充吞吐量 (tok/s) |
|---|---|---|---|---|
| 1 | step-3.7-flash-q3km_20260607T061506Z | 3588.9 | 71.2 | 1454.0 |
| 2 | step-3.7-flash-iq4xs_20260607T044220Z | 3493.8 | 81.4 | 1830.6 |
| 3 | step-3.7-flash-q3km_20260607T140502Z | 3194.6 | 38.1 | 1338.3 |
| 4 | step-3.7-flash-iq4xs_20260607T135829Z | 1850.0 | 53.8 | 1626.0 |
| 5 | mimo-v2-flash-iq2m_20260607T141330Z | 1917.9 | 63.2 | 1080.6 |
| 6 | mimo-v2-flash-iq2xxs_20260607T140641Z | 1841.0 | 58.4 | 1213.3 |
| 7 | mimo-v2-flash-iq2m_20260607T054242Z | 2122.0 | 85.2 | 1154.2 |
运行轴键
当前唯一的活跃矩阵轴是 model × benchmark。其他轴(如 agent、engine、judge、GPU、vLLM 版本)保持恒定,并作为一级键记录在每个运行的 full_run_manifest.json 中。
| 排名 | 运行标识 | agent | engine | judge | GPU | vLLM |
|---|---|---|---|---|---|---|
| 1 | qwopus3.6-27b-v2_... | hermes | vllm | gpt-5.4 | NVIDIA B300 SXM6 AC | 0.22.0 |
| 2 | nemotron-ultra-3-mopd-... | hermes | vllm | gpt-5.4 | NVIDIA B300 SXM6 AC | 0.22.0 |
| 3 | exo-qwen3.6-27b-nvfp4_... | hermes | vllm | gpt-5.4 | NVIDIA B300 SXM6 AC | — |
| 4 | qwen3.6-27b-nvfp4_... | hermes | vllm | gpt-5.4 | NVIDIA B300 SXM6 AC | 0.22.0 |
| 5 | qwen3.6-35b-a3b_... | hermes | vllm | gpt-5.4 | NVIDIA B300 SXM6 AC | 0.22.0 |
| 6 | step-3.7-flash-q3km_... | hermes | vllm | gpt-5.4 | NVIDIA B200 | — |
| 7 | qwen3.6-35b-a3b-awq-4bit_... | hermes | vllm | gpt-5.4 | NVIDIA B300 SXM6 AC | 0.22.0 |
| 8 | step-3.7-flash-iq4xs_... | hermes | vllm | gpt-5.4 | NVIDIA B200 | — |
| 9 | qwen3.6-27b-text-nvfp4-mtp_... | hermes | vllm | gpt-5.4 | NVIDIA B300 SXM6 AC | 0.22.0 |
| 10 | qwen3.6-27b-fp8_... | hermes | vllm | gpt-5.4 | NVIDIA B200 | 0.22.0 |
| 11 | gemma-4-31b-it_... | hermes | vllm | gpt-5.4 | NVIDIA B200 | 0.22.0 |
| 12 | qwen3.6-35b-a3b-fp8_... | hermes | vllm | gpt-5.4 | NVIDIA B300 SXM6 AC | 0.22.0 |
| 13 | gemma-4-31b-it-awq-4bit_... | hermes | vllm | gpt-5.4 | NVIDIA B300 SXM6 AC | 0.22.0 |
| 14 | step-3.7-flash-q3km_... | hermes | vllm | gpt-5.4 | NVIDIA B200 | — |
| 15 | qwen3.6-27b_... | hermes | vllm | gpt-5.4 | NVIDIA B300 SXM6 AC | 0.22.0 |
| 16 | step-3.7-flash-iq4xs_... | hermes | vllm | gpt-5.4 | NVIDIA B200 | — |
| 17 | gpt-oss-120b_... | hermes | vllm | gpt-5.4 | NVIDIA B300 SXM6 AC | 0.22.0 |
| 18 | gemma-4-26b-a4b-it-awq-4bit_... | hermes | vllm | gpt-5.4 | NVIDIA B200 | — |




