env0-experiment-trajectories
收藏资源简介:
Env0 Experiment Trajectories 是一个私有数据集,旨在保存来自 2026 年 6 月 16-17 日 Env0 后训练试点的完整 BenchFlow 评估轨迹工件。数据集内容主要包括:位于 `eval_summaries/` 目录下的 BenchFlow 任务级别摘要、每个任务的结果与奖励元数据,以及从基础和适配器评估运行中捕获的 ACP 轨迹文件。此外,还包含 2026 年 6 月 19 日 SkillsBench 10 任务管道金丝雀测试的完整工件同步,涵盖工件审计报告和每次 rollout 的轨迹。数据集在上传前已扫描常见实时令牌模式。由于轨迹中可能仍包含本地路径、模拟服务状态、提供者请求元数据以及其他不应未经单独审查即公开的实验上下文信息,因此该数据集被设为私有。这些工件的目的是为了长期保存和确保可审计性,它们记录了 BenchFlow -> Modal vLLM -> Pi-ACP -> 验证器路径产生了带有奖励的评估轨迹,但其本身并不直接证明模型性能的提升。
Env0 Experiment Trajectories is a private dataset designed to preserve complete BenchFlow evaluation trajectory artifacts from the Env0 post-training pilot conducted on June 16–17, 2026. The dataset primarily comprises: BenchFlow task-level summaries stored in the `eval_summaries/` directory, result and reward metadata for each task, as well as ACP trajectory files captured from both base model and adapter evaluation runs. Additionally, it includes full artifact synchronization from the SkillsBench 10-task pipeline canary test held on June 19, 2026, covering artifact audit reports and trajectories from each rollout. The dataset was scanned for common live token patterns prior to upload. Given that the trajectories may still contain local paths, simulated service states, provider request metadata, and other experimental contextual information that should not be disclosed without separate review, this dataset is designated as private. The purpose of these artifacts is long-term preservation and auditability; they document reward-bearing evaluation trajectories generated via the BenchFlow -> Modal vLLM -> Pi-ACP -> Validator pipeline, but do not directly demonstrate improved model performance.
数据集概述
名称:Env0 Experiment Trajectories
语言:英语
标签:benchflow, env0, trajectories, tool-use
可见性:私有数据集
内容
该数据集保存了2026年6月16日至17日Env0后训练试点项目中的完整BenchFlow评估轨迹工件。数据来自经净化的会话导出源仓库:https://github.com/benchflow-ai/env0-experiment
数据集包含以下文件结构:
eval_summaries/**/summary.json:BenchFlow任务级摘要eval_summaries/**/result.json:每个任务的结果和奖励元数据eval_summaries/**/trajectory/acp_trajectory.jsonl:基础和适配器评估运行中捕获的ACP轨迹experiments/skillsbench/qwen35-skillsbench-sglang/canary10-20260619-150811/**:2026年6月19日SkillsBench 10任务管道金丝雀的完整工件同步,包括工件审计报告和每个部署轨迹
上传文件在提交前已扫描常见实时令牌模式。数据集为私有,因为轨迹中可能仍包含本地路径、模拟服务状态、提供者请求元数据及其他未经过单独审查不应公开的实验上下文信息。
范围
这些工件用于长期保存和可审计性。它们记录了BenchFlow → Modal vLLM → Pi-ACP → 验证器路径产生了可获奖励的评估轨迹,但数据集本身并不单独证明模型性能提升。





