Experiments
Collect MLflow experiments and runs by project, compare runs, and inspect a run's metrics, parameters and artifacts.
Required permission: View
This screen collects the record (run) each training leaves. Runs from platform training and runs logged to MLflow from your own PC all land here if they belong to the same tenant. ML engineers use it to compare results and register a good run in the model registry.
Records are grouped as project → experiment → run. The project comes from an MLflow experiment tag; experiments without the tag are grouped under default.
Experiment list
| # | Area | Description |
|---|---|---|
| ① | Search | Filter by experiment name or run id (run names are not searched). It searches only the runs already loaded on the screen. |
| ② | Compare (비교) | Enabled when two or more runs are checked. The number in parentheses is how many you selected. |
| ③ | Experiment | Experiment name and number of runs. Each experiment first shows its 20 most recent runs; if there are more, Load more runs (Run 더 보기) loads 20 more at a time. The trash icon on the right deletes the experiment (develop, after a confirmation). |
| ④ | Compare selection | Pick a run for comparison. You can pick runs from different experiments. |
| ⑤ | Run | Name (short id if unnamed), status, headline metric, start time. The headline metric is one whose name contains acc, iou, f1 or score, otherwise the first metric. Click the name to open the run detail. Runs created by platform training get a Training (학습) link to the training run detail. |
Compare runs
Put two or more runs side by side. The Metrics (메트릭) tab opens first, and rows whose values differ are highlighted in yellow.
| # | Area | Description |
|---|---|---|
| ① | Tabs | Overview (개요) (status, start time), Metrics (메트릭), Parameters (파라미터), Charts (그래프). |
| ② | Show differences only (차이만 보기) | On the Metrics (메트릭) and Parameters (파라미터) tabs, hides rows whose value is the same in every run. A value present in only some runs counts as different. |
| ③ | Comparison table | The first column is the metric or parameter name; the other columns are runs (short id). |
The Charts (그래프) tab overlays the runs' curves on one chart per metric. The enlarge icon at the top right of a chart shows it larger.
Run detail
| # | Area | Description |
|---|---|---|
| ① | Summary | Run name and status, experiment ID, start and end, duration, number of metrics. Runs created by platform training get a View training run (학습 실행 보기) link. |
| ② | Actions | Register model (모델 등록) (model_register, FINISHED runs only), Stop (중지) (develop, runs that have not ended, after a confirmation), Delete (삭제) (develop, finished, failed or killed runs only). Without the permission the button is hidden; in the wrong state it is greyed out. |
| ③ | Tabs | Overview (개요) (final metrics, parameters), Metrics (메트릭) (charts), Parameters (파라미터) (search), Tags (태그), Artifacts (아티팩트) (file browser). |
The Metrics (메트릭) tab shows every logged metric as a small chart (the all (전체) view). Click a chart to switch to the single (단일) view, which also shows the last, min, max, mean and number of steps; change the metric with the select box at the top right. With only one metric, the single view opens directly.
The Artifacts (아티팩트) tab shows the files the run uploaded (weights, result images, logs) as a tree. Select a file to see a preview on the right (image, text, JSON, first 200 rows of a CSV) and download it with Download (다운로드). Text and CSV files over 2 MB and binaries such as weights have no preview, only download. Download all (전체 다운로드) gets everything as a zip.
Main tasks
- In the experiment list, check two or more runs to compare and press Compare (비교).
Experiments — ① search ② Compare (two or more selected) ③ experiment ④ select a run to compare ⑤ run name (opens detail) - On the Metrics (메트릭) tab, use Show differences only (차이만 보기) to see only the metrics that differ.
Compare runs — ① Overview, Metrics, Parameters and Charts tabs ② show differences only ③ values per run (differences highlighted) - Open the best run and press Register model (모델 등록). Enter a model name and press Register (등록); the new version goes into Staging, and View version (버전 보기) in the confirmation message opens it. Registering again under the same name increments the version number.
Register to the model registry — the new version enters Staging
Model registration looks for the model file in the run's artifacts in the order model logged with MLflow → model/ folder → weights/best.pt → weights/last.pt, and uses the first one found. A version registered from a single file under weights/ has no MLflow model metadata (MLmodel), so no serving image can be built from it.
Statuses and badges
| Badge | Value | Meaning |
|---|---|---|
| Running (실행 중) | RUNNING | Training is in progress and metrics are accumulating. |
| Finished (완료) | FINISHED | Ended normally. Models can be registered only in this state. |
| Failed (실패) | FAILED | Ended with an error. |
| Killed (중단됨) | KILLED | Stopped by a user, or ended because platform training was stopped or interrupted. |
| Scheduled (예약됨) | SCHEDULED | Not started yet. |