caiovicentino1/swebench-pro-qwen36-27b-phase6
收藏资源简介:
该数据集名为SWE-bench Pro Qwen3.6-27B Phase 6 — Trajectories + Residuals,是Tool-Entropy Collapse论文和Two Honest Nulls论文(进行中)的配套数据。它包含99个多轮次智能体轨迹,这些轨迹来自Qwen3.6-27B模型在SWE-bench Pro任务(涉及qutebrowser、openlibrary、ansible)上的运行结果,并捕获了每轮次在L11、L23、L31、L43、L55层的残差流激活(以bf16格式的safetensors存储)。此外,数据集还提供了所有衍生特征和标签,用于复现论文并进行下游分析。主要用途包括:检查评估工具中的tool_entropy_collapse评估、对残差流进行机制解释研究、分析SWE-bench Pro的失败模式,以及作为智能体监控的基准(提供WANDERING/SUCCESS/LOCKED标签)。数据集结构包括轨迹文件、捕获的激活数据、衍生特征文件等。来源方面,模型为Qwen3.6-27B,任务为SWE-bench Pro,运行于NVIDIA RTX 6000 Pro Blackwell硬件,日期为2026年5月。数据集还包含关于运行稳定性的注意事项,特别是WANDERING类别的标签可能受单次运行分类和温度采样随机性的影响。
This dataset is named SWE-bench Pro Qwen3.6-27B Phase 6 — Trajectories + Residuals, serving as supporting data for the papers *Tool-Entropy Collapse* and *Two Honest Nulls* (in progress). It contains 99 multi-turn AI Agent trajectories generated by the Qwen3.6-27B model during its execution on SWE-bench Pro tasks (covering qutebrowser, openlibrary, and ansible), and captures residual stream activations at layers L11, L23, L31, L43, and L55 for each turn, stored in bf16-formatted safetensors files. Additionally, the dataset provides all derived features and labels for paper reproduction and downstream analysis. Its core use cases include: evaluating the tool_entropy_collapse metric in inspection tools, conducting mechanistic interpretability research on residual streams, analyzing failure modes of SWE-bench Pro, and acting as a benchmark for agent monitoring with labels of WANDERING/SUCCESS/LOCKED. The dataset structure consists of trajectory files, captured activation data, derived feature files, and other relevant assets. For the dataset source: the model used is Qwen3.6-27B, the tasks are SWE-bench Pro, the experiments were conducted on NVIDIA RTX 6000 Pro Blackwell hardware, and the generation date is May 2026. The dataset also includes notes on runtime stability, specifically that the labels for the WANDERING category may be affected by single-run classification and temperature sampling randomness.




