open-llm-leaderboard/details_TFLai__gpt2-turkish-uncased
收藏Hugging Face2023-12-02 更新2024-03-04 收录
下载链接：
https://hf-mirror.com/datasets/open-llm-leaderboard/details_TFLai__gpt2-turkish-uncased
下载链接
链接失效反馈官方服务：
资源简介：
该数据集是在评估模型TFLai/gpt2-turkish-uncased时自动创建的，由64个配置组成，每个配置对应一个评估任务。数据集由3次运行生成，每次运行的结果作为特定分割存储在配置中，分割名称为运行的时间戳。train分割始终指向最新结果。此外，还有一个名为results的配置，存储了所有运行的聚合结果，并用于计算和显示在Open LLM Leaderboard上的聚合指标。
This dataset was automatically created during the evaluation of the model TFLai/gpt2-turkish-uncased. It comprises 64 configurations, each corresponding to one evaluation task. The dataset is generated from three runs, with the results of each run stored as a dedicated split within its corresponding configuration, where the split name is the timestamp of the corresponding run. The 'train' split always points to the most recent result. Additionally, there exists a configuration named 'results' that stores the aggregated results across all runs, and is utilized to compute and display the aggregate metrics on the Open LLM Leaderboard.
提供机构：
open-llm-leaderboard原始信息汇总
数据集概述
该数据集是在评估模型 TFLai/gpt2-turkish-uncased 在 Open LLM Leaderboard 上的运行过程中自动创建的。
数据集组成
- 数据集包含 64 个配置，每个配置对应一个评估任务。
- 数据集从 3 次运行中创建，每次运行可以在每个配置中找到特定的分割，分割名称使用运行的时间戳。
- "train" 分割始终指向最新的结果。
- 一个额外的配置 "results" 存储所有运行的聚合结果，用于计算和显示 Open LLM Leaderboard 上的聚合指标。
数据加载示例
python from datasets import load_dataset data = load_dataset("open-llm-leaderboard/details_TFLai__gpt2-turkish-uncased", "harness_gsm8k_5", split="train")
最新结果
以下是 2023-12-02T15:29:40.186292 运行的最新结果：
python { "all": { "acc": 0.0, "acc_stderr": 0.0 }, "harness|gsm8k|5": { "acc": 0.0, "acc_stderr": 0.0 } }
配置详情
-
config_name: harness_arc_challenge_25
- split: 2023_07_24T09_48_46.264649
- path: **/details_harness|arc:challenge|25_2023-07-24T09:48:46.264649.parquet
- split: latest
- path: **/details_harness|arc:challenge|25_2023-07-24T09:48:46.264649.parquet
- split: 2023_07_24T09_48_46.264649
-
config_name: harness_drop_3
- split: 2023_10_22T01_34_05.823968
- path: **/details_harness|drop|3_2023-10-22T01-34-05.823968.parquet
- split: latest
- path: **/details_harness|drop|3_2023-10-22T01-34-05.823968.parquet
- split: 2023_10_22T01_34_05.823968
-
config_name: harness_gsm8k_5
- split: 2023_10_22T01_34_05.823968
- path: **/details_harness|gsm8k|5_2023-10-22T01-34-05.823968.parquet
- split: 2023_12_02T15_29_40.186292
- path: **/details_harness|gsm8k|5_2023-12-02T15-29-40.186292.parquet
- split: latest
- path: **/details_harness|gsm8k|5_2023-12-02T15-29-40.186292.parquet
- split: 2023_10_22T01_34_05.823968
-
config_name: harness_hellaswag_10
- split: 2023_07_24T09_48_46.264649
- path: **/details_harness|hellaswag|10_2023-07-24T09:48:46.264649.parquet
- split: latest
- path: **/details_harness|hellaswag|10_2023-07-24T09:48:46.264649.parquet
- split: 2023_07_24T09_48_46.264649
-
config_name: harness_hendrycksTest_5
- split: 2023_07_24T09_48_46.264649
- path:
- **/details_harness|hendrycksTest-abstract_algebra|5_2023-07-24T09:48:46.264649.parquet
- **/details_harness|hendrycksTest-anatomy|5_2023-07-24T09:48:46.264649.parquet
- **/details_harness|hendrycksTest-astronomy|5_2023-07-24T09:48:46.264649.parquet
- **/details_harness|hendrycksTest-business_ethics|5_2023-07-24T09:48:46.264649.parquet
- **/details_harness|hendrycksTest-clinical_knowledge|5_2023-07-24T09:48:46.264649.parquet
- **/details_harness|hendrycksTest-college_biology|5_2023-07-24T09:48:46.264649.parquet
- **/details_harness|hendrycksTest-college_chemistry|5_2023-07-24T09:48:46.264649.parquet
- **/details_harness|hendrycksTest-college_computer_science|5_2023-07-24T09:48:46.264649.parquet
- **/details_harness|hendrycksTest-college_mathematics|5_2023-07-24T09:48:46.264649.parquet
- **/details_harness|hendrycksTest-college_medicine|5_2023-07-24T09:48:46.264649.parquet
- **/details_harness|hendrycksTest-college_physics|5_2023-07-24T09:48:46.264649.parquet
- **/details_harness|hendrycksTest-computer_security|5_2023-07-24T09:48:46.264649.parquet
- **/details_harness|hendrycksTest-conceptual_physics|5_2023-07-24T09:48:46.264649.parquet
- **/details_harness|hendrycksTest-econometrics|5_2023-07-24T09:48:46.264649.parquet
- **/details_harness|hendrycksTest-electrical_engineering|5_2023-07-24T09:48:46.264649.parquet
- **/details_harness|hendrycksTest-elementary_mathematics|5_2023-07-24T09:48:46.264649.parquet
- **/details_harness|hendrycksTest-formal_logic|5_2023-07-24T09:48:46.264649.parquet
- **/details_harness|hendrycksTest-global_facts|5_2023-07-24T09:48:46.264649.parquet
- **/details_harness|hendrycksTest-high_school_biology|5_2023-07-24T09:48:46.264649.parquet
- **/details_harness|hendrycksTest-high_school_chemistry|5_2023-07-24T09:48:46.264649.parquet
- **/details_harness|hendrycksTest-high_school_computer_science|5_2023-07-24T09:48:46.264649.parquet
- **/details_harness|hendrycksTest-high_school_european_history|5_2023-07-24T09:48:46.264649.parquet
- **/details_harness|hendrycksTest-high_school_geography|5_2023-07-24T09:48:46.264649.parquet
- **/details_harness|hendrycksTest-high_school_government_and_politics|5_2023-07-24T09:48:46.264649.parquet
- **/details_harness|hendrycksTest-high_school_macroeconomics|5_2023-07-24T09:48:46.264649.parquet
- **/details_harness|hendrycksTest-high_school_mathematics|5_2023-07-24T09:48:46.264649.parquet
- **/details_harness|hendrycksTest-high_school_microeconomics|5_2023-07-24T09:48:46.264649.parquet
- **/details_harness|hendrycksTest-high_school_physics|5_2023-07-24T09:48:46.264649.parquet
- **/details_harness|hendrycksTest-high_school_psychology|5_2023-07-24T09:48:46.264649.parquet
- **/details_harness|hendrycksTest-high_school_statistics|5_2023-07-24T09:48:46.264649.parquet
- **/details_harness|hendrycksTest-high_school_us_history|5_2023-07-24T09:48:46.264649.parquet
- **/details_harness|hendrycksTest-high_school_world_history|5_2023-07-24T09:48:46.264649.parquet
- **/details_harness|hendrycksTest-human_aging|5_2023-07-24T09:48:46.264649.parquet
- **/details_harness|hendrycksTest-human_sexuality|5_2023-07-24T09:48:46.264649.parquet
- **/details_harness|hendrycksTest-international_law|5_2023-07-24T09:48:46.264649.parquet
- **/details_harness|hendrycksTest-jurisprudence|5_2023-07-24T09:48:46.264649.parquet
- **/details_harness|hendrycksTest-logical_fallacies|5_2023-07-24T09:48:46.264649.parquet
- **/details_harness|hendrycksTest-machine_learning|5_2023-07-24T09:48:46.264649.parquet
- **/details_harness|hendrycksTest-management|5_2023-07-24T09:48:46.264649.parquet
- **/details_harness|hendrycksTest-marketing|5_2023-07-24T09:48:46.264649.parquet
- **/details_harness|hendrycksTest-medical_genetics|5_2023-07-24T09:48:46.264649.parquet
- **/details_harness|hendrycksTest-miscellaneous|5_2023-07-24T09:48:46.264649.parquet
- **/details_harness|hendrycksTest-moral_disputes|5_2023-07-24T09:48:46.264649.parquet
- **/details_harness|hendrycksTest-moral_scenarios|5_2023-07-24T09:48:46.264649.parquet
- **/details_harness|hendrycksTest-nutrition|5_2023-07-24T09:48:46.264649.parquet
- **/details_harness|hendrycksTest-philosophy|5_2023-07-24T09:48:46.264649.parquet
- **/details_harness|hendrycksTest-prehistory|5_2023-07-24T09:48:46.264649.parquet
- **/details_harness|hendrycksTest-professional_accounting|5_2023-07-24T09:48:46.264649.parquet
- **/details_harness|hendrycksTest-professional_law|5_2023-07-24T09:48:46.264649.parquet
- **/details_harness|hendrycksTest-professional_medicine|5_2023-07-24T09:48:46.264649.parquet
- **/details_harness|hendrycksTest-professional_psychology|5_2023-07-24T09:48:46.264649.parquet
- **/details_harness|hendrycksTest-public_relations|5_2023-07-24T09:48:46.264649.parquet
- **/details_harness|hendrycksTest-security_studies|5_2023-07-24T09:48:46.264649.parquet
- **/details_harness|hendrycksTest-sociology|5_2023-07-24T09:48:46.264649.parquet
- **/details_harness|hendrycksTest-us_foreign_policy|5_2023-07-24T09:48:46.264649.parquet
- path:
- split: 2023_07_24T09_48_46.264649
搜集汇总
数据集介绍
构建方式
在大型语言模型评估领域，该数据集是专为模型TFLai/gpt2-turkish-uncased在Open LLM Leaderboard平台上的评测运行而自动生成的。其构建过程涉及64个独立配置，每个配置精确对应一项被评估的任务。数据来源于三次独立的运行，每次运行的结果均以时间戳为标识，作为特定分割（split）存储于各配置中。特别地，名为“train”的分割始终指向最新一次运行的成果。此外，一个名为“results”的附加配置汇集了所有运行的综合结果，用于在排行榜上计算和展示聚合指标。
特点
该数据集的核心特色在于其结构化与动态性。它通过多配置设计，将不同任务的评估细节分门别类，涵盖了从常识推理到数学问题求解的广泛领域，如ARC挑战、GSM8K及HellaSwag等。每次独立运行的时间戳分割，忠实记录了模型性能的演变轨迹，使得纵向对比成为可能。同时，“latest”分割始终指向最新数据，确保了研究者总能获取最前沿的评估结果。这种设计兼顾了历史追溯性与实时更新性，为深入分析模型能力提供了坚实的数据基础。
使用方法
研究者可通过HuggingFace的datasets库便捷地调用该数据集。例如，使用load_dataset函数，指定数据集名称“open-llm-leaderboard/details_TFLai__gpt2-turkish-uncased”，并选择目标配置（如“harness_gsm8k_5”）及所需分割（如“latest”或具体时间戳），即可加载对应任务的详细评估数据。若需访问聚合结果，则选用“results”配置。这一流程简洁高效，使得研究者能够轻松聚焦于特定任务或时间点的模型表现，进行深入的性能剖析与比较分析。
背景与挑战
背景概述
在大规模语言模型快速演进的浪潮中，如何系统、公正地评估模型在多样化任务上的表现成为研究焦点。Open LLM Leaderboard由Hugging Face团队于2023年创建，旨在为开源大语言模型提供一个标准化、透明化的竞技平台。该数据集是围绕土耳其语GPT-2模型（TFLai/gpt2-turkish-uncased）的评估运行而自动生成的，记录了模型在ARC挑战、HellaSwag、GSM8K、DROP以及涵盖57个学科的MMLU等多项基准测试中的详细结果。通过存储多次运行的时间戳分片与聚合指标，该数据集为研究者提供了复现、对比和分析模型能力的可靠依据，推动了多语言、低资源场景下语言模型评估的规范化进程。
当前挑战
该数据集所解决的领域问题在于，多语言尤其是低资源语言（如土耳其语）的大语言模型缺乏统一、细粒度的评估体系，现有基准多聚焦于英语，导致模型在非英语语境下的真实能力难以度量。构建过程中面临的挑战包括：一是评估任务的多样性导致数据配置复杂，需为64个任务分别维护独立的分片结构，且每次运行的时间戳分片增加了数据管理的难度；二是模型在GSM8K等推理任务上准确率极低（如记录显示为0.0），暴露出小规模模型在复杂推理能力上的根本性不足，这对评估指标的敏感性和区分度提出了更高要求。
常用场景
经典使用场景
在自然语言处理领域，对于低资源语言如土耳其语的模型评估长期面临基准匮乏的困境。该数据集围绕土耳其语GPT-2模型的评测而生，覆盖了从常识推理（如HellaSwag）、数学求解（如GSM8K）到多学科知识（如MMLU）的多样化任务，为研究者提供了一个标准化、可复现的评估框架。其核心价值在于通过细粒度的任务配置，支持对模型在特定能力维度上的深入剖析，从而揭示模型在土耳其语语境下的优势与局限。
衍生相关工作
该数据集作为Open LLM Leaderboard生态的一部分，衍生了一系列关于模型评测与开源基准建设的重要工作。它启发了针对土耳其语及其他低资源语言的模型评估标准化研究，推动了如ELIYA等土耳其语评测集的构建。同时，基于其多任务配置，研究者开展了关于模型规模、训练数据量与下游性能之间关系的实证分析，为更高效的模型训练策略提供了参考，也促进了跨语言模型能力对比的学术讨论。
数据集最近研究
最新研究方向
在非英语自然语言处理领域，特别是针对低资源语言如土耳其语的预训练语言模型评估正成为前沿热点。该数据集记录了gpt2-turkish-uncased模型在Open LLM Leaderboard上的多任务评测结果，涵盖ARC挑战、HellaSwag、GSM8K及涵盖57个学科的MMLU基准测试，揭示了当前土耳其语GPT-2模型在复杂推理与知识密集型任务上表现薄弱的现状。这一评测体系不仅为土耳其语大语言模型的性能基准提供了标准化参照，更推动了多语言模型评估框架的完善，其意义在于促进语言技术向更广泛语种的普惠发展，并为后续针对土耳其语的模型优化与数据增强策略指明了关键瓶颈。
以上内容由遇见数据集搜集并总结生成
