AgentSkillsCN

run-single-trial

当仅需运行一次端到端实验时，务必确保环境已正确搭建，仅运行一个随机种子，并将输出存储于可复现的运行目录中。

SKILL.md
--- frontmatter
name: run-single-trial
description: Use when running a single end-to-end experiment trial. Ensure the environment is set up, run exactly one seed, and store outputs in a reproducible run directory.

Run exactly one experiment trial end-to-end.

Steps:

  1. •Confirm environment installation steps exist (env/ + README).
  2. •Choose a canonical command (e.g., python -m scripts.train --config configs/paper.yaml --seed 0).
  3. •Run the command and capture:
    • •stdout/stderr
    • •results/<run_id>/config.json
    • •results/<run_id>/metrics.*
    • •results/<run_id>/summary.json
  4. •Validate basic sanity checks:
    • •learning curve is not all NaN
    • •episodic return is in plausible bounds
    • •eval runs produce outputs
  5. •Update README with the one-liner command and expected runtime.