The Advanced step of a new AutoML job is optional. Each section has defaults that suit most jobs; expand one to change it. Which sections appear depends on the problem type:
| Section | Tabular | Multimodal | Time Series |
|---|
| Model Selection | Yes | | |
| Ensemble & Stacking | Yes | | |
| Data Handling & Features | Yes | | |
| Training Behavior | Yes | Verbosity only | Verbosity only |
| Hyperparameter Optimization (HPO) | Yes | | |
| Model Hyperparameters | Yes | | |
| Time Series Configuration | | | Yes |
| Multimodal Configuration | | Yes | |
| Custom JSON Configuration | Yes | | |
Model Selection
The model families to train. The header counts the enabled ones. Fewer families train faster.
| Group | Family | On by default |
|---|
| Gradient Boosting | LightGBM (GBM): fast gradient boosting | Yes |
| CatBoost (CAT): strong on categorical features | Yes |
| XGBoost (XGB): robust gradient boosting | Yes |
| Tree-Based | Random Forest (RF): stable, interpretable | Yes |
| Extra Trees (XT): a faster random forest variant | Yes |
| Neural Networks | PyTorch Neural Network (NN_TORCH) | Yes |
| FastAI Neural Network (FASTAI) | Yes |
| Transformer (FT_TRANSFORMER): GPU recommended | No |
| Other Models | K-Nearest Neighbors (KNN) | Yes |
| Linear/Logistic Regression (LR) | Yes |
| TabPFN: a foundation model for up to 10,000 rows | No |
Ensemble & Stacking
Bagging
| Setting | Default | Meaning |
|---|
| Bag Folds (K-fold) | 0 (off) | Train each model on K folds and average them. 5 to 10 is typical; 0 to 20. |
| Bag Sets (Repeats) | 1 | Repeat the bagging with new folds. More sets are better but slower; 1 to 20. |
| Save bag fold models | On | Keep each fold's model |
Stacking
| Setting | Default | Meaning |
|---|
| Stack Levels | No stacking | 1, 2 or 3 levels (3 is slow) of models trained on the previous level's predictions |
| Dynamic stacking | Off | Detects stacking overfitting and decides the levels itself |
| Fit weighted ensemble in each layer | On | A weighted ensemble of each layer's models |
| Include original features in stacker | On | Stacked models also see the original columns |
Post-Training
| Setting | Default | Meaning |
|---|
| Refit best model on full data | Off | Refit the best model on the training and validation rows together |
| Set refitted as best model | Off | Serve the refitted model |
| Keep only best model | Off | Delete the other models to save space |
| Delete auxiliary files | Off | Delete files not needed for prediction |
Data Handling & Features
| Setting | Default | Meaning |
|---|
| Feature Generator | Auto (recommended) | AutoGluon's automatic feature engineering (dates, categories, text). Disabled uses the columns as they are. |
| Sample Weight Column | None | A column whose values weight each row |
| Generate text n-gram features | On | Word n-gram counts from text columns |
| Extract text special features | On | Lengths, symbol and word counts from text columns |
| Include raw text as features | Off | Pass text through as text |
| Enable feature pruning | Off | Drop features that do not help the score |
Training Behavior
| Setting | Default | Meaning |
|---|
| Inference Time Limit (seconds/row) | 0 (no limit) | The slowest prediction per row allowed; slower models are left out |
| Fit Strategy | Sequential (stable) | Parallel fits models at the same time with Ray (experimental) |
| Verbosity Level | Normal | Silent, Minimal, Normal, Detailed or Debug: how much the training log says |
| Calibrate probabilities (classification) | Auto | Temperature-scale a classifier's probabilities on the validation rows |
| Calibrate decision threshold (binary) | Auto | Choose the threshold that maximizes the metric |
| Track learning curves | Off | Record each model's metric over its iterations |
| Raise on model failure | Off | Fail the job when one model fails, instead of skipping it (debugging) |
Hyperparameter Optimization (HPO)
Off by default: every family trains once per bag with its defaults. Enable hyperparameter tuning searches each family's hyperparameters over AutoGluon's search space for it.
| Setting | Default | Choices |
|---|
| Search Strategy | Auto | Auto (Bayesian for neural networks, random for the others), Random Search, Bayesian Optimization |
| Number of Trials | 10 | 1 to 100, per family |
| Time Limit per Trial (seconds) | 300 | 60 to 3,600 |
| Scheduler | Local | Local or FIFO |
LightGBM, CatBoost, XGBoost, the linear model and the neural networks are tuned. Random Forest, Extra Trees and K-Nearest Neighbors have no search space and train once with their defaults. The job's time limit still applies: tuning stops when it runs out.
Model Hyperparameters
Off by default: every family trains with AutoGluon's defaults for the preset, and tuning searches them. Turn on Customize model hyperparameters to fix values for every model of a family; tuning then searches only the hyperparameters you did not set.
| Tab | Hyperparameters (starting values) |
|---|
| LightGBM | Boosting Rounds 1000, Num Leaves 128, Learning Rate 0.05, Feature Fraction 0.9, Min Data in Leaf 20, Extra Trees mode off |
| CatBoost | Iterations 1000, Tree Depth 6, Learning Rate 0.05, L2 Regularization 3, Random Strength 1 |
| XGBoost | Estimators 1000, Max Depth 6, Learning Rate 0.05, Subsample 0.9, Colsample by Tree 0.9, Min Child Weight 1 |
| Random Forest | Estimators 300, Max Features sqrt (or log2), Min Samples Leaf 1 |
| Neural Net | Epochs 100, Learning Rate 0.001, Dropout 0.1, Batch Size 256 (32 to 1,024), Hidden Layers 3, Hidden Size 256 (64 to 1,024) |
Time Series Configuration
| Setting | Default | Meaning |
|---|
| Prediction Length (Horizon) | 1 | How many steps ahead to forecast (1 to 365) |
| Frequency | Daily (D) | Minute (T), Hourly (H), Daily (D), Weekly (W), Monthly (M), Quarterly (Q) or Yearly (Y) |
| ID Column (Series Identifier) | | The column naming each series, such as item_id or store_id |
| Timestamp Column | | The column holding each row's time |
| Known Covariates | | Comma-separated columns whose future values are known when forecasting, such as holidays or promotions |
A forecaster produces a point forecast and the 0.1, 0.5 and 0.9 quantiles.
Multimodal Configuration
| Setting | Default | Meaning |
|---|
| Text Columns | | Comma-separated columns read as free text |
| Image Columns | | Comma-separated columns naming image files in the dataset's volume; the job fetches them before training |
| Text Backbone | BERT Base | BERT Base, RoBERTa Base, ELECTRA Base or DistilBERT |
| Max Text Length (tokens) | 512 | 128, 256 or 512 |
| Image Size (pixels) | 224x224 | 224x224, 384x384 or 512x512 |
| Enable image augmentation | On | Randomly augments training images (AutoGluon's trivial augment) |
| Enable multimodal ensembling | Off | Train several multimodal models, one per backbone, and ensemble them |
Custom JSON Configuration
Tabular jobs only, for advanced users. Each setting must be a JSON object; Next refuses one that is not, naming it and why, and so does the REST API.
- Custom Hyperparameters (JSON): AutoGluon's
hyperparameters argument, keyed by family, for example {"GBM": {"num_boost_round": 500}, "CAT": {"iterations": 500}}. A family named here replaces that family's settings from Model Hyperparameters.
- ag_args_fit (JSON): arguments for every model's fit, for example
{"num_cpus": 4, "num_gpus": 1}.
The REST API and SDK also take ag_args_json and ag_args_ensemble_json (see the AutoML REST API). Switching a job to Multimodal or Time Series clears its custom JSON.