name: ztx-exp-manager description: Manage ML/LLM experiment workflows with ztxexp. Use when the agent needs to build parameter configurations, run experiments, validate run artifacts, analyze results, or troubleshoot failed runs.
ztx-exp-manager Skill(中文)
核心规则
- 优先使用
ExperimentPipeline组织实验,不手写零散调度脚本。 - 实验函数契约固定为:
exp_fn(ctx: RunContext) -> dict | None。 - 成功判定必须使用:
run.json.status == "succeeded"。 - 产物分工:
- 最终指标:
return dict->metrics.json - 过程指标:
ctx.log_metric(...)->metrics.jsonl - 业务文件:写入
artifacts/
- 最终指标:
常用流程
- 先用
grid/variants/random_search构建配置。 - 再执行
pipeline.run(exp_fn, mode=...)。 - 用
ResultAnalyzer聚合、导出和清理。
常用命令
pip install -U ztxexpztxexp init-vibepython -m pytest
排障清单
- 先看
run.json.status。 - 失败时查看
error.log。 - 曲线缺失时检查
ctx.log_metric(...)是否调用。
ztx-exp-manager Skill (English)
Core Rules
- Prefer
ExperimentPipelinefor workflow orchestration. - Keep experiment contract:
exp_fn(ctx: RunContext) -> dict | None. - Determine success strictly by
run.json.status == "succeeded". - Artifact responsibilities:
- Final metrics:
return dict->metrics.json - Step metrics:
ctx.log_metric(...)->metrics.jsonl - Business outputs: write into
artifacts/
- Final metrics:
Typical Workflow
- Build configs via
grid/variants/random_search. - Execute with
pipeline.run(exp_fn, mode=...). - Aggregate and clean results using
ResultAnalyzer.
Common Commands
pip install -U ztxexpztxexp init-vibepython -m pytest
Troubleshooting
- Check
run.json.statusfirst. - Read
error.logon failure. - Verify
ctx.log_metric(...)for missing curves.