Jev-Agent-Trace-Audit-100K: a dataset for auditing software agent runs on Hugging Face
MaziyarPanahi · x · 2026-11-06
Developer MaziyarPanahi released Jev-Agent-Trace-Audit-100K on Hugging Face: models read a software agent's full run trace, then decide how the run ended, which tools were used, and whether a correct patch was written.
It's part of a larger Jev release — 5 datasets spanning 1.8B tokens read and 720,000 answered decisions, covering tool calls, agent traces, customer support, routing, and policy checks, with source labels and answers on every row so you can train your own tiny in-house decider. Intended uses include trace auditing, failure analysis, and reward shaping; no clinical data in any of the 5 datasets.
Related event: Jev Releases 720K Decision Data for Training Lightweight Routing Models(3 posts)→
More from coding & agent
- Cursor launches live in-editor app building with one-link sharing, from idea to deployed app — mattyp · 2026-11-06
- Claude Code Web gotcha: headless Chrome has no GPU, so WebGL silently fell back to CPU — julianharris · 2026-11-06
- 69 alternatives to Conductor mapped as coding-agent orchestrators compete — DavidWells · 2026-11-06
- Sentry CEO shares screenshot: one AI coding agent spent $2,867 in 7 days — zeeg · 2026-11-06
- LangChain shares mental models for high-quality LLM evals from harmonic's Austin Berke — LangChain · 2026-11-06
- GraphRail: graph-grounded step-level verification catches search agent reasoning errors without an LLM judge — _reachsumit · 2026-11-06