Examples

Example scripts demonstrating openadapt-ml usage.

Training from JSON Data

The standard workflow for using openadapt-ml:

from openadapt_ml.ingest import load_episodes
from openadapt_ml.schemas import validate_episodes, summarize_episodes

# 1. Load your exported data
episodes = load_episodes("your_data/")

# 2. Validate against schema
warnings = validate_episodes(episodes, check_images=True)

# 3. Check summary
print(summarize_episodes(episodes))

Full Training Example

# Validate data only
python examples/train_from_json.py --data your_data/ --validate-only

# Train a model
python examples/train_from_json.py \
  --data your_data/ \
  --output results/ \
  --config configs/qwen3vl_capture.yaml

Data Format

Episodes should be JSON files following the openadapt-ml schema. See sample_data.json for the expected format.

Required Fields

Episode:

• id (string): Unique identifier
• goal (string): Task description
• steps (array): List of Step objects

Step:

• t (float): Timestamp in seconds
• observation (object): GUI state
• action (object): Action taken

Observation:

• image_path (string): Path to screenshot

Action:

• type (string): One of: click, double_click, right_click, drag, scroll, type, key, done
• x, y (float, 0-1): Normalized coordinates (for click/drag)
• text (string): Text content (for type action)
• key (string): Key name (for key action)

Optional Fields

See openadapt_ml/schemas/sessions.py for full schema with all optional fields.