hf-hub-session-captures
收藏资源简介：
该数据集名为dacorvo/hf-hub-session-captures，是一个记录代理（agent）与模型（model）之间交互过程的HTTP捕获数据集。每条数据记录（对应数据表的一行）代表一次对/v1/chat/completions API端点的调用，数据由HuggingFace的开源工具agentcap生成。该数据集是hf-hub-session集合的一部分，与名为hf-hub-session-<agent>-traces的原始会话追踪数据集是配套关系，两者可以通过共享的`run_id`字段进行关联和联合分析。数据集包含11个字段，详细记录了每次交互的元数据和内容。核心字段包括：`run_id`（用于关联追踪数据集的代理运行标识符）、`request_id`（由捕获代理生成的唯一请求ID）、`model`（请求中指定的模型标识符）、`captured_at`（请求被捕获的时间戳，单位为纪元秒）。最关键的内容字段是`request`和`response`，它们分别以JSON字符串的形式存储了原始的OpenAI格式请求体和响应体。对于流式响应（Server-Sent Events），`response`字段有特定的结构（`{"stream": true, "raw": ...}`）。此外，数据集还包含了与HuggingFace路由服务相关的头部信息字段：`served_by`（X-Served-By响应头，指示路由的子提供商）、`served_build_info`（X-Build-Info响应头）、`served_model`（响应体回显的模型标识）。最后，`provider`和`upstream_url`字段分别记录了捕获时代理上游的提供商标识和具体URL。使用该数据集时，需要将`request`和`response`字段进行JSON解析。如需恢复请求中每条消息对应的token范围，需要用户自行使用相应模型的tokenizer（如`transformers.AutoTokenizer.apply_chat_template`）对请求消息进行渲染。该数据集适用于分析代理-模型的交互模式、研究API调用行为、进行模型响应质量评估或路由策略分析等任务。
The dataset named dacorvo/hf-hub-session-captures is an HTTP capture dataset that records the interaction process between agents and models. Each data record (corresponding to a row in the data table) represents a call to the /v1/chat/completions API endpoint, and the data is generated by HuggingFaces open-source tool agentcap. This dataset is part of the hf-hub-session collection and is complementary to the original session trace dataset named hf-hub-session-<agent>-traces, with both being linkable via the shared `run_id` field for joint analysis. The dataset contains 11 fields that detail the metadata and content of each interaction. Core fields include: `run_id` (agent run identifier for linking with trace datasets), `request_id` (unique request ID generated by the capture agent), `model` (model identifier specified in the request), `captured_at` (timestamp when the request was captured, in epoch seconds). The most critical content fields are `request` and `response`, which store the original OpenAI format request body and response body as JSON strings, respectively. For streaming responses (Server-Sent Events), the `response` field has a specific structure (`{"stream": true, "raw": ...}`). Additionally, the dataset includes header-related fields associated with HuggingFace routing services: `served_by` (X-Served-By response header indicating the sub-provider of the route), `served_build_info` (X-Build-Info response header), `served_model` (model identifier echoed in the response body). Finally, the `provider` and `upstream_url` fields record the provider identifier and specific URL of the agents upstream at the time of capture. When using this dataset, the `request` and `response` fields require JSON parsing. To recover the token ranges corresponding to each message in the request, users need to render the request messages using the appropriate models tokenizer (e.g., `transformers.AutoTokenizer.apply_chat_template`). This dataset is suitable for tasks such as analyzing agent-model interaction patterns, studying API call behaviors, evaluating model response quality, or analyzing routing strategies.
数据集概述
数据集名称：dacorvo/hf-hub-session-captures
许可证：Apache-2.0
数据集标签：agentcap, agentcap-captures
数据集描述：该数据集包含了代理（agent）与模型（model）之间交互的HTTP捕获记录，每一行对应一次 /v1/chat/completions 调用。数据集由 agentcap 工具生成。
配套数据集：相同的运行会话的本地跟踪记录存储在同名的 hf-hub-session-<agent>-traces 数据集中，所有数据集均归属于 hf-hub-session Collection，通过 run_id 字段进行关联。
加载方式
python from datasets import load_dataset
ds = load_dataset("dacorvo/hf-hub-session-captures", split="train")
数据模式（Schema）
| 列名 | 描述 |
|---|---|
run_id |
agentcap 运行ID；与跟踪数据集中每个运行文件夹匹配 |
request_id |
由捕获代理生成的UUID |
model |
捕获请求体中的模型ID |
captured_at |
请求被捕获时的Unix时间戳（秒） |
request |
原始OpenAI请求体，JSON字符串化 |
response |
原始OpenAI响应体，JSON字符串化（对于SSE流式响应，格式为 {"stream": true, "raw": ...}） |
served_by |
每个响应中的 X-Served-By 头部（HF Router子提供商路由） |
served_build_info |
每个响应中的 X-Build-Info 头部 |
served_model |
响应体中回显的 model 字段 |
provider |
从代理上游URL推导得出（每个文件恒定） |
upstream_url |
捕获时的代理上游URL（每个文件恒定） |
数据处理说明
request和response字段为JSON字符串，使用时需通过json.loads()解析。- 若要恢复每条消息的token范围，需自行将
request.messages通过模型的聊天模板进行渲染，可使用transformers.AutoTokenizer.apply_chat_template方法。
