Create a Workflow

Before You Start

A training project already created in Datature Vi
A dataset with annotations available in your organization
A clear idea of your VLM task (phrase grounding or visual question answering)

A workflow in Datature Vi is a saved training configuration. It captures three things: the system prompt that instructs the model, the dataset and split settings, and the model architecture and hyperparameters. You can run the same workflow multiple times and get consistent, reproducible results.

Open the workflow canvas

Inside your training project, click Create Workflow. The canvas opens with two nodes arranged top to bottom: System Prompt and Dataset.

You should see

The workflow appears in your training project's Overview page under the Workflows section. You can now configure your workflow and launch a training run.

Do this with the Vi SDK

import vi

client = vi.Client(
    secret_key="your-secret-key",
    organization_id="your-organization-id"
)

flow = client.flows.create(
    spec={
        "name": "My Training Workflow",
        "schema": "workspaces/datature/flowSchemas/default",
        "tolerations": {},
        "settings": {
            "execution": {
                "accelerator": {"name": "NvidiaA10G", "count": 1}
            },
            "features": {"evaluationPreview": True},
            "evaluationIntervalEpochs": 1
        },
        "blocks": [
            {
                "block": "workspaces/datature/blocks/system-prompt",
                "settings": {
                    "systemPrompt": "Your system prompt here."
                },
                "style": {
                    "type": "system-prompt",
                    "position": {"x": 0, "y": -412}
                }
            },
            {
                "block": "workspaces/datature/blocks/dataset",
                "settings": {
                    "splitStrategy": {
                        "strategy": "TrainTestSplit",
                        "splitRatio": 0.2,
                        "shuffle": True,
                        "seed": 0
                    },
                    "project": "your-dataset-project",
                    "projectKind": "phrase-grounding"
                },
                "style": {
                    "type": "dataset",
                    "position": {"x": 0, "y": -231}
                }
            },
            {
                "block": "workspaces/datature/blocks/qwen3vl-model",
                "settings": {
                    "architecture": {"name": "qwen3vl", "size": "8B"},
                    "trainingMode": {"type": "LoRA"},
                    "quantization": {"enabled": True, "type": "NF4"},
                    "compute": {"precisionType": "BFloat16"},
                    "hyperparameters": {
                        "epochs": 10,
                        "learningRate": 0.0001,
                        "batchSize": 8,
                        "gradientAccumulation": 1,
                        "optimizer": "AdamW"
                    },
                    "evaluation": {
                        "maxNewTokens": 512,
                        "topK": 50,
                        "topP": 1,
                        "temperature": 1,
                        "repetitionPenalty": 1.05
                    }
                },
                "style": {
                    "type": "qwen3vl-model",
                    "position": {"x": 0, "y": 0}
                }
            }
        ],
        "trainingProject": "workspaces/your-org/trainingProjects/your-project"
    }
)
print(f"Created: {flow.spec.name} ({flow.flow_id})")

For more details, see the full SDK reference.

Next steps

Start A Training Run

Configure training settings, select GPU hardware, and start a training run.

Configure Your System Prompt

Learn how to write effective prompts for phrase grounding, VQA, and custom tasks.

Configure Your Model

Compare model architectures and tune hyperparameters for your use case.