pinchbench-clawd
收藏资源简介:
PinchBench Clawd训练数据是一个合成微调数据集,旨在训练大型语言模型(LLM)作为OpenClaw框架中的自主AI代理Clawd。数据集针对PinchBench基准测试(包含23个任务),每个样本均为多轮对话,展示Clawd如何利用文件I/O、网络搜索、电子邮件、日历、图像生成、记忆等多种工具完成现实世界任务。数据通过Anthropic Batch API生成,并经过LLM法官评分(1-5分)和质量过滤。数据集包含训练集(train.jsonl,评分≥3/5且去重)、验证集(val.jsonl)和评分反馈文件(scores.json)。生成流程采用自动化代理循环(包括评估代理、诊断代理、数据生成代理、数据策展代理和训练代理)。该数据集采用CC BY 4.0许可协议发布。
PinchBench Clawd training data is a synthetic fine-tuning dataset aimed at training large language models (LLMs) to serve as the autonomous AI agent Clawd within the OpenClaw framework. This dataset is tailored for the PinchBench benchmark, which encompasses 23 tasks. Each sample in the dataset is a multi-turn dialogue that illustrates how Clawd leverages a variety of tools including file I/O, web search, email, calendar, image generation, and memory to accomplish real-world tasks. The dataset was generated using the Anthropic Batch API, and subjected to scoring by LLM judges (1-5 points) and quality filtering. It consists of three parts: the training set (train.jsonl, with scores ≥3/5 and deduplicated), the validation set (val.jsonl), and a score feedback file (scores.json). The dataset's generation pipeline follows an automated agent loop, which includes an evaluation agent, diagnostic agent, data generation agent, data curation agent, and training agent. This dataset is released under the CC BY 4.0 license.
PinchBench Clawd Training Data 数据集概述
基本信息
- 数据集名称:PinchBench Clawd Training Data
- 许可证:CC BY 4.0
- 任务类别:文本生成
- 语言:英语
- 标签:合成数据、微调、工具使用、智能体、OpenClaw、PinchBench
- 数据集格式:JSONL、JSON
数据集描述
该数据集是一个合成的微调数据集,用于训练大型语言模型(LLM)充当 Clawd——一个在 OpenClaw 框架上的自主AI智能体。其目标针对 PinchBench 基准测试(包含23个任务)。
每个示例都是一个多轮对话,其中Clawd使用各种工具(如文件I/O、网络搜索、电子邮件、日历、图像生成、记忆等)来完成一项现实世界任务。数据通过Anthropic Batch API使用Claude生成,由LLM评判员进行评分(1-5分),并经过质量过滤。
文件构成
- train.jsonl:训练示例(评分 >= 3/5,已去重)
- val.jsonl:保留的验证示例
- scores.json:每个示例的LLM评判员评分和反馈
生成流程
数据通过一个智能体循环(pinchbenchmaxing)生成:
- EvalAgent:在PinchBench上对模型进行基准测试
- EvalAnalysisAgent:使用Claude诊断失败原因
- DataAgent:基于诊断结果生成有针对性的训练数据
- CuratorAgent:对数据进行评分、修复边界示例、去重,并推送至此
- TrainerAgent:进行微调,然后循环重复
使用许可
本数据集根据 CC BY 4.0 许可证发布。您可以出于任何目的自由分享和改编数据,但需注明出处。
引用信息
如果使用本数据集,请引用:
@misc{pinchbench-clawd-2026, author = {Chetan Tekur}, title = {PinchBench Clawd Training Data}, year = {2026}, url = {https://huggingface.co/datasets/cptekur/pinchbench-clawd}, }




