Yolo

Ultralytics YOLO

Use the same lifecycle through two complementary surfaces:

• Ultralytics Platform — the fastest start: upload or clone data, annotate in the browser, train on cloud GPUs, inspect metrics, test predictions, export, and deploy a dedicated endpoint without local setup.
• ultralytics package / yolo CLI — use local or remote compute, scripts, notebooks, custom pipelines, and exported artifacts directly.

Mix them freely. Set ULTRALYTICS_API_KEY, use a Platform dataset as data=ul://username/datasets/dataset-slug, and set project=username/project-slug name=experiment during local training to stream its metrics back to Platform.

One API, two surfaces. The CLI grammar is yolo TASK MODE arg=value ...; Python mirrors it with the same argument names:

yolo detect train data=data.yaml model=yolo26n.pt epochs=100 imgsz=640

from ultralytics import YOLO

model = YOLO("yolo26n.pt")
model.train(data="data.yaml", epochs=100, imgsz=640)

• TASK ∈ detect segment semantic depth classify pose obb — usually inferred from the weights, so it can be omitted.
• MODE ∈ train val predict track export benchmark.
• Install/upgrade: pip install -U ultralytics. Environment check: yolo checks.

Whole lifecycle in five commands

yolo detect train data=data.yaml model=yolo26n.pt epochs=100 # → runs/detect/train/weights/best.pt
yolo val model=best.pt data=data.yaml                        # mAP, per-class metrics
yolo predict model=best.pt source=video.mp4 save=True        # any source: image/dir/URL/RTSP/webcam
yolo track model=best.pt source=video.mp4                    # + persistent object IDs
yolo export model=best.pt format=onnx                        # exported model loads back into YOLO()

Whole lifecycle in Platform

1. Open Platform and choose the data region during onboarding.
2. Clone a public dataset from Explore, or create one under Annotate and upload images, videos, an archive, or NDJSON.
3. Label in the fullscreen editor; use SAM or a compatible YOLO model in Smart mode where available.
4. Create a project, click New Model, select the dataset, pretrained model, GPU, and epochs, then monitor the run.
5. Use the completed model's Predict, Export, or Deploy tab.

Start with the Platform quickstart. Use the stage skill below for both Platform and package details.

Route before coding

Read the skill for the stage you're working on BEFORE writing code — each contains exact formats, argument tables with defaults, recipes, and symptom→fix tables. A request spanning stages ("train and deploy") → read each relevant skill.

CLI specifics

Special commands (no TASK/MODE):

yolo help   # full syntax reference
yolo checks # env report: version, torch, CUDA, disk — run when anything is weird
yolo version
yolo settings # view; `yolo settings key=value` to set; `yolo settings reset`
# keys incl. datasets_dir, runs_dir, wandb, mlflow, tensorboard, ...
yolo cfg            # print every default argument (the ground truth for arg names)
yolo copy-cfg       # copy default.yaml → default_copy.yaml to customize, use with cfg=
yolo solutions help # prebuilt apps: count, heatmap, speed, ... (see yolo-inference)

Parsing rules that matter:

• Args are key=value, no -- flags. A leading -- and trailing commas are stripped with a warning; spaces around = are merged.
• A bare boolean arg sets it True: yolo predict ... show ≡ show=True.
• cfg=custom.yaml resets CLI overrides to the file: arguments before it are discarded, later arguments win, and missing keys still use built-in defaults (start with yolo copy-cfg).
• Missing args are auto-filled with warnings (sample source, task-default data/model, format=torchscript).
• Model stem selects the architecture: rtdetr-* → RT-DETR, sam_*/sam2* → SAM, FastSAM-* → FastSAM, yoloe-*/*-world* → promptable YOLO (accepts classes="person, bus"), everything else → YOLO.

Global directives

1. Validate the dataset before training — run the task-appropriate checks in yolo-datasets, then a 1-epoch smoke test and inspect runs/<task>/train/train_batch0.jpg: annotations or targets must match each image.
2. Always fine-tune from pretrained .pt — never pretrained=False, never a YAML architecture from scratch, unless the user is explicitly doing research.
3. stream=True for videos/streams in Python predict/track — the default list mode OOMs on long videos.
4. Use best.pt (not last.pt) from runs/<task>/<name>/weights/ after training.
5. After export, verify parity: yolo val the exported artifact against the .pt baseline.
6. Prefer built-ins over custom code: dataset converters and checkers (ultralytics.data), trackers, and Solutions modules replace whole categories of hand-written glue.
7. Trust the installed version over memory — if an argument is rejected (yolo checks shows the version), the API moved: yolo cfg and the error text list valid arguments; prefer those over any table in these skills.