Jevotron：從命令列進行多項 JEV 整合

Jevotron：從命令列進行多項 JEV 整合
Jevotron: Multiple Jev integrations from the command line

原始链接: https://cmungall.github.io/jevotron/

Jevotron 使用选定字段、自然语言指导、可选示例和可配置的警告阈值，扫描结构化文件中的异常。工作流分为三个阶段： 1. **预览**——在消耗 API 调用之前，检查数据块、字段路径以及实际的模型请求。 2. **扫描**——统一评估每个条目，并保留字段级概率、条目分数、来源位置和日期。 3. **审查**——优先处理可疑条目，将结果导出为 CSV，或将 JSONL 数据传入现有的 Shell 工作流。支持 CSV、YAML、JSON、TOML、文本、OBO 和压缩文件。通过字段选择、指导文件和本地 Python 配置，可以实现可复用或自定义的解析。SQLite 用于持久化成功结果、缓存未更改的条目，并支持重新排序、调整阈值以及恢复中断的运行，而无需重新评估未更改的数据。应用示例包括检测错误的机场所属国家、审查库存、评估本体定义以及对智能体执行轨迹进行分类。在一项包含 24 条轨迹的试点测试中，Jevotron 匹配了 163 个步骤质量标签中的 130 个（79.8%），有害步骤检测的精确率为 89.7%，召回率为 70.3%。

``` Hacker News 最新 | 往期 | 评论 | 提问 | 展示 | 工作 | 提交登录 Jevotron：从命令行进行多种 Jev 集成 ( cmungall.github.io ) 12 分作者： chrismungall 23 小时前 | 隐藏 | 往期 | 收藏 | 1 条评论 help octoberfranklin 21 小时前 [–] 我们能不能别再搞这种愚蠢的机器人水军把戏了？回复指南 | 常见问题 | 列表 | API | 安全 | 法律 | 申请加入 YC | 联系我们搜索： ```

jevotron / airports.csv

$ jevotron scan airports.csv --guidance "Check airport locations."

01 / PREVIEW

See what goes in

Inspect chunks, field paths, and exact model requests before making an API call.

Preview a file →

02 / SCAN

Assess every entry

Jev scores selected fields together. Each entry gets the same guidance and optional examples.

Choose your fields →

03 / REVIEW

Start with the warnings

Sort suspicious entries, export CSV, or pipe JSONL into your existing shell workflow.

Build a review queue →

Small setup. Useful defaults.

CSV, YAML, JSON, TOML, text, OBO, and more work out of the box, including gzip files. The format reference covers defaults and format-specific options. Select fields with --field, add a sentence with --guidance, and run. Longer instructions can come from --guidance-file. A local Python config is available when a project needs custom parsing or reusable settings.

Unchanged input reuses its assessment. SQLite saves each successful result as it arrives. Reorder a file, change a reporting threshold, or resume a failed run without reassessing unchanged entries.

A review aid with visible evidence. Reports retain each field's probabilities, the entry score, source location, and assessment date. The warning score is the highest field anomaly probability; you choose the threshold.

Try a complete example

Example	What you'll do
Airports / CSV	Find two injected country errors in public data, then compare versions.
Inventory / YAML	Apply written rules and a chosen exemplar to stock records.
Measurement units / OBO	Score definitions and repeated synonyms within independent stanzas.
Agent traces / JSONL	Classify public agent traces and individual steps, then compare with published labels.

Agent traces: a measured pilot

On a small, length-filtered sample of 24 public traces, jt matched 130 of 163 step-quality labels (79.8%). Harmful-step precision was 89.7%, with 70.3% recall. The example uses original messages and tool definitions, with human labels withheld from the model.

Try the trace example → · Read the full analysis and limitations →