semianalysisai/cc-traces-weka-042026

CC Traces — Weka (April 2026)数据集是一个包含739个多轮代理跟踪（约59.3k个独立模型请求）的集合，由`claude-opus-4-5-20251101`驱动。每个跟踪记录了单个代理会话的完整请求/响应序列，包括每个请求的KV块哈希，使得数据集可以用于重放推理引擎或模拟离线前缀缓存行为。数据集还提供了详细的统计信息和图表，展示了输入序列长度、输出序列长度、客户端思考时间等指标的分布情况。该数据集适用于推理基准测试、KV缓存研究和容量规划等用途。

CC Traces — Weka (April 2026) is a collection of 739 multi-turn agentic traces (≈ 59.3k individual model requests) driven by `claude-opus-4-5-20251101`. Each trace captures the full request/response sequence of a single agent session, including per-request KV block hashes, so the dataset can be replayed against an inference engine or used to simulate prefix-cache behavior offline. The dataset includes detailed statistics and plots showing the distribution of metrics such as input sequence length, output sequence length, and client think time. It is intended for uses such as inference benchmarking, KV-cache research, and capacity planning.