ppb-results
收藏资源简介:
Poor Paul's Benchmark (PPB) 是一个用于在消费级、专业级和小型企业硬件上进行本地AI推理的开源基准测试框架。该数据集包含社区提交的跨模型、量化、硬件、运行时和基准测试设置的标准化基准测试结果。每一行代表一个基准测试结果,旨在用于开放基准测试、可重复性和下游分析。常见用例包括比较硬件间的推理吞吐量、研究上下文长度扩展、比较延迟指标(如TTFT和ITL)以及支持仪表板和派生排行榜。数据集结构为每行对应一个标准化基准测试结果,字段按重要性排列,包括模型和基准测试身份、硬件、基准测试设置、性能指标、系统元数据和提交及来源字段。数据集目前使用单一分割(train)。数据来源为用户在本地硬件上执行的PPB运行,经过标准化处理后上传。使用该数据时需注意其作为追加式原始提交日志的局限性,包括结果来自异构系统、可能受热力学和驱动程序等因素影响、某些指标仅适用于特定运行器类型等。
Poor Paul's Benchmark (PPB) is an open-source benchmarking framework for local AI inference on consumer-grade, professional, and small business hardware. This dataset contains standardized benchmark results submitted by the community across various models, quantization settings, hardware platforms, runtimes, and benchmark configurations. Each row represents a single benchmark result, intended for open benchmarking, reproducibility, and downstream analysis. Common use cases include comparing inference throughput across different hardware, investigating context length expansion, comparing latency metrics such as TTFT and ITL, and supporting dashboard development and derived leaderboards. The dataset is structured such that each row corresponds to a standardized benchmark result, with fields ordered by importance, including model and benchmark identifiers, hardware details, benchmark setup parameters, performance metrics, system metadata, as well as submission and source fields. Currently, the dataset uses a single split: "train". The data originates from PPB runs executed by users on their local hardware, which are standardized and then uploaded. When using this dataset, attention should be paid to its limitations as an append-only raw submission log, including that results come from heterogeneous systems, may be affected by factors such as thermodynamics and device drivers, and that certain metrics only apply to specific runner types.
数据集概述:Poor Pauls Benchmark Results
数据集简介
Poor Pauls Benchmark (PPB) 是一个用于在消费级、专业消费级和小型企业硬件上进行本地AI推理的开源基准测试框架。本数据集包含社区提交的、经过标准化的基准测试结果,涵盖模型、量化、硬件、运行时和基准测试设置。每一行代表一个基准测试结果,旨在用于开放的基准测试、可重复性研究以及下游分析。
数据集结构
数据集采用扁平化的表格结构,每一行对应一个标准化的基准测试结果。字段分组如下:
- 模型与基准测试标识:
model、model_base、quant、runner_type - 硬件信息:
gpu_name、gpu_vram_gb、gpu_driver、backends、cpu_model - 基准测试设置:
n_ctx、n_batch、concurrent_users - 性能指标:
throughput_tok_s、avg_ttft_ms、p50_ttft_ms、p99_ttft_ms、avg_itl_ms、p50_itl_ms、p99_itl_ms - 系统元数据:
os_system、os_release、os_machine、cpu_cores、ram_total_gb - 提交与溯源信息:
submitter、timestamp、submitted_at、schema_version、benchmark_version、submission_id、row_id、machine_fingerprint、run_fingerprint、result_fingerprint、source_file_sha256
部分字段是特定于运行器的,因此可能存在null值。
数据创建
源数据来自用户在其本地硬件上执行的PPB运行。在上传前,原始基准测试输出被标准化为扁平化的表格模式,以便在Hugging Face上预览,并方便pandas、DuckDB、电子表格和下游仪表板使用。
使用注意事项
本数据集是一个仅追加的原始提交账本,而非最终经过筛选的排行榜。
重要限制包括:
- 结果来自异构的真实世界系统。
- 散热、驱动程序、后台负载和本地调优可能影响结果。
- 部分指标仅适用于特定的运行器类型。
- 按设计可能存在重复和重复提交。
- 不包含成本相关指标。
对于下游分析:
- 使用
result_fingerprint识别完全重复的行。 - 使用
run_fingerprint对相同基准测试标识的重复运行进行分组。 - 使用
machine_fingerprint对来自同一匿名机器的结果进行分组。
附加信息
- 许可证:MIT
- 项目仓库:https://github.com/paulplee/poor-pauls-benchmark
- 数据集仓库:https://huggingface.co/datasets/paulplee/ppb-results
- 标签:benchmarking, llama-cpp, llm-inference, local-llm, homelab, open-data, tabular




