everydayness-agent-traces
收藏资源简介:
Everydayness — Anonymized Agent Traces 是一个包含匿名化执行轨迹的数据集,这些轨迹来源于为 Build Small Hackathon(2026年6月由 Gradio 与 Hugging Face 联合举办)所开发的Everydayness悲伤陪伴应用程序。数据集旨在通过公开分享匿名的管道执行轨迹,帮助社区了解小规模模型(如 Qwen3-14B、Whisper)在真实悲伤陪伴分类任务上的实际表现,并因此获得了该黑客马拉松的Sharing is Caring徽章。数据以 JSON Lines 格式存储于 anonymized_traces.jsonl 文件中,每行代表一次管道执行轨迹,数据规模小于1000条样本。每条轨迹记录包含以下匿名化字段:会话ID(UUID v4,不关联用户信息)、运行的管道类型(pain_filter、audio、media、session_summary之一)、使用的模型名称及量化信息、消息分块处理的批次大小、分类标签分布比例(不包含消息内容)、推理延迟毫秒数、使用的GPU类型以及提示词模板摘要(仅包含占位符描述)。数据集严格不记录任何个人信息,包括消息内容、发送者姓名、聊天时间戳、任何个人身份信息以及文件内容或转录文本。该数据集适用于文本分类任务的研究与分析,特别是关注小模型在情感支持或对话代理场景下性能评估的社区。相关应用程序信息可在 Hugging Face Space(build-small-hackathon/everydayness)查看,其开发属于 Backyard AI 赛道,并使用了 OpenPipe/Qwen3-14B-Instruct(AWQ 4-bit量化)模型。数据集语言涉及英语、泰卢固语和印地语,采用 MIT 许可证。
Everydayness — Anonymized Agent Traces is a dataset containing anonymized execution traces, sourced from the Everydayness grief companion application developed for the Build Small Hackathon (jointly hosted by Gradio and Hugging Face in June 2026). The dataset aims to help the community understand the real-world performance of small-scale models (such as Qwen3-14B and Whisper) in grief companion classification tasks by publicly sharing anonymized pipeline execution traces, and it has earned the Sharing is Caring badge from the hackathon. The data is stored in JSON Lines format in the anonymized_traces.jsonl file, with each line representing a pipeline execution trace, and the dataset size is less than 1000 samples. Each trace record includes anonymized fields: session ID (UUID v4, not linked to user information), pipeline type (one of pain_filter, audio, media, session_summary), model name and quantization information, batch size for message chunking, classification label distribution ratios (excluding message content), inference latency in milliseconds, GPU type used, and prompt template summary (only placeholder descriptions). The dataset strictly does not record any personal information, including message content, sender names, chat timestamps, any personally identifiable information, or file content or transcription text. It is suitable for research and analysis of text classification tasks, particularly for communities focusing on performance evaluation of small models in emotional support or dialogue agent scenarios. Related application information can be viewed on Hugging Face Space (build-small-hackathon/everydayness), developed as part of the Backyard AI track, using the OpenPipe/Qwen3-14B-Instruct (AWQ 4-bit quantized) model. The dataset involves languages such as English, Telugu, and Hindi, and is licensed under MIT.
数据集名称
Everydayness — Anonymized Agent Traces
许可证
MIT
任务类别
文本分类(text-classification)
语言
英语(en)、泰卢固语(te)、印地语(hi)
标签
everydayness、grief-companion、llm-traces、hackathon、build-small-hackathon
数据集规模
少于1000条(n<1K)
数据集内容
数据集包含一个名为 anonymized_traces.jsonl 的文件,其中每条记录对应一次流水线执行轨迹。
记录的字段包括:
session_id:UUID v4,不含任何用户相关信息pipeline:执行的流水线名称(如 pain_filter、audio、media、session_summary)model:模型名称及量化方式batch_size:每个批次处理的消息数classification_distribution:标签分布比例(不含消息内容)inference_latency_ms:Modal调用的推理延迟(毫秒)gpu:使用的GPU类型prompt_template_summary:模板描述(包含占位符)
绝对不包含的内容:
- 消息内容
- 发送者姓名
- 聊天时间戳
- 任何个人身份信息
- 文件内容或转录文本
数据集用途
该数据集用于公开分享匿名的流水线执行轨迹,帮助社区理解小模型(如 Qwen3-14B、Whisper)在真实哀伤陪伴分类任务上的表现,并用于获取 Build Small Hackathon 的 📡 Sharing is Caring 徽章。
相关应用与背景
- HF Space: build-small-hackathon/everydayness
- 赛道: Backyard AI
- 模型: OpenPipe/Qwen3-14B-Instruct (AWQ 4-bit, xgrammar, Modal A100)





