메뉴
Query SwanLab experiment metrics and run status. Use when checking training/producer metrics, comparing runs, or investigating experiment health. Keywords: swanlab, metrics, experiment, loss, reward, training progress, wandb.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
Launch a background agent to monitor a long-running remote job (K8s, SLURM, CI). Use when submitting jobs that need async monitoring. Keywords: monitor, watch job, background agent, submit and watch, check job status.
Generate a weekly report summarizing commits and PRs across the main repo and all external/ submodules. Keywords: weekly report, 周报, work summary, 工作总结.
Capture reusable knowledge after milestones. Systematically review what was learned and persist it to the right layer. Keywords: sediment, crystallize, capture, persist, remember, learned, milestone, retrospective.
Hypothesis-first bug diagnosis. Verify the user's diagnosis before proposing alternatives. Keywords: bug, diagnose, debug, fix, hypothesis, root cause, crash, error, timeout, traceback.
Build a first-failure timeline for distributed incidents. Use when debugging crashes, hangs, watchdog timeouts, collective mismatches, or cascading errors where the last error is often not the root cause.
Convert local debugging results into high-quality upstream issues or RFC comments. Use when filing bugs to external projects with clear reproduction, first-failure evidence, root-cause hypothesis, and scoped fix suggestions.
| name | swanlab |
| description | Query SwanLab experiment metrics and run status. Use when checking training/producer metrics, comparing runs, or investigating experiment health. Keywords: swanlab, metrics, experiment, loss, reward, training progress, wandb. |
Query experiment metrics from SwanLab. All commands run via pixi run -e dev python3 -c "...".
Already configured. Default workspace: junjayzhang, project: dev.
import swanlab
api = swanlab.Api()
# All running experiments
for r in api.runs(path='junjayzhang/dev', filters={'state': 'RUNNING'}):
print(f'{r.name} | {r.id} | {r.state} | {r.created_at}')
# Filter by state: RUNNING, FINISHED, CRASHED, ABORTED
run = api.run(path='junjayzhang/dev/<experiment_id>')
df = run.metrics(keys=['<metric_key>'], sample=50) # returns pandas DataFrame
Producer runs (*_producer):
group/reward_mean, group/reward_min, group/reward_maxgroup/n_truncated, group/truncated_ratiogroup/n_non_truncated, group/success_ratiogroup/length_mean, group/length_min, group/length_maxgroup/output_tokens_mean, group/output_tokens_maxgroup/python_calls_mean, group/python_errors_totalgroup/n_context_exhausted, group/n_max_turnsTrain runs (*_train):
loss_metrics/global_avg_loss, loss_metrics/global_max_losstime_metrics/end_to_end(s), time_metrics/data_loading(s), time_metrics/data_loading(%)optimization/learning_rate (or optimization/scheduler_0/learning_rate)run.name # str
run.id # experiment_id
run.state # RUNNING | FINISHED | CRASHED | ABORTED
run.created_at # ISO 8601
run.url # web UI link
When asked to check experiment metrics or training progress, use this skill to:
run.metrics(keys=[...], sample=N)Always wrap API calls in pixi run -e dev python3 -c "..." since swanlab is only in the dev environment.