Re-training Models¶
After the agent completes a run, you can re-train the model with new data using the train.sh script.
When to Use¶
- Train on updated or expanded datasets
- Fine-tune with additional samples
- Reproduce training with different data splits
Basic Usage¶
./train.sh \
--agent-dir outputs/<agent_id> \
--train-data /path/to/new_train.csv \
--validation-data /path/to/new_validation.csv \
--artifacts-dir /path/to/output_artifacts
Required Arguments¶
| Argument | Description |
|---|---|
--agent-dir |
Path to completed agent output folder |
--train-data |
Path to new training CSV file |
--validation-data |
Path to new validation CSV file |
--artifacts-dir |
Where to save new training artifacts |
Optional Arguments¶
| Argument | Description |
|---|---|
--cpu-only |
Run without GPU |
--local |
Run locally without Docker |
--help |
Show help message |
Example¶
# Re-train using new data
./train.sh \
--agent-dir outputs/enchanted_fixing_reigned \
--train-data datasets/updated_data/train.csv \
--validation-data datasets/updated_data/validation.csv \
--artifacts-dir outputs/retrained_model
How It Works¶
The script:
- Loads the agent's
train.pyscript frombest_run_files/ - Uses the agent's conda environment
- Runs training with the new data
- Saves artifacts to the specified directory
Data Format¶
Your new data files must match the format expected by the agent's training script:
- Same column names as original training data
- Same feature encoding/preprocessing expectations
- Target column with same name and format
Output¶
After training completes, you'll find:
Docker vs Local Mode¶
By default, training runs in Docker for isolation. Use --local for direct execution:
# Docker mode (default)
./train.sh --agent-dir outputs/my_agent ...
# Local mode
./train.sh --local --agent-dir outputs/my_agent ...
GPU Support¶
GPU is used automatically if available. To disable:
Troubleshooting¶
"Docker image not found"¶
Run ./run.sh once to build the Docker image, or use --local mode.
"Agent directory not found"¶
Ensure the path points to a completed agent output in outputs/.
"Column mismatch"¶
Your new data must have the same column structure as the original training data.
Next Steps¶
- Running Inference - Make predictions with trained models
- Understanding Outputs - Explore what the agent produces