Skip to main content

​
ExperimentConfig

The main configuration class that orchestrates all aspects of model training.
​
DataConfig | List[DataConfig]
required
Data configuration(s) for potentially multi-objective modeling.
​
OptimizationConfig
required
Optimization and training parameters.
​
ModelConfig
required
Model architecture and initialization parameters.
​
MetaConfig
default:"Uses MetaConfig defaults"
Metadata and run-specific parameters.

​
MetaConfig

Configuration for experiment metadata and checkpointing behavior.
​
str
default:"trial-run"
Name identifier for this experimental run.
​
int
default:"42"
Random seed for reproducible training.
​
str
default:"current working directory / run_name"
Directory path for saving model checkpoints.
​
int
default:"-1"
Frequency (in steps) for saving model checkpoints. Set to -1 to save only at the end of training.
​
int
default:"-1"
Maximum number of model checkpoints to retain. Set to -1 for no limit.
​
WandbConfig | None
default:"None"
Weights & Biases logging configuration for experiment tracking and visualization.

​
WandbConfig

Configuration for Weights & Biases experiment tracking and logging.
​
str
required
Weights & Biases project name for organizing experiments.
​
str | None
default:"None"
Weights & Biases team/organization name. If None, uses the default entity associated with your API key.
​
str | None
default:"None"
Custom run name for the experiment. If None, uses the MetaConfig name or auto-generates one.
​
List[str] | None
default:"None"
List of tags to associate with the run for easy filtering and organization.
​
str | None
default:"None"
Optional notes or description for the experiment run.
​
bool
default:"False"
Whether to log the model as a Weights & Biases artifact for version control.
​
int
default:"100"
Frequency (in steps) for logging metrics to Weights & Biases.
​
bool
default:"False"
Whether to log gradient histograms (can impact performance).
​
bool
default:"False"
Whether to log parameter histograms (can impact performance).
​
str | None
default:"None"
Model watching mode for logging gradients and parameters:
  • "gradients": Log gradient histograms
  • "parameters": Log parameter histograms
  • "all": Log both gradients and parameters
  • None: Disable model watching
​
dict | None
default:"None"
Additional configuration dictionary to log to Weights & Biases.

​
DataConfig

Configuration for training data and objectives. Can be specified as a single instance or list for multi-task learning.
​
str | List[str]
required
Path(s) to preprocessed data files.
​
str | List[str] | callable | List[callable] | None
default:"None"
Feature engineering functions for lag tokens (historical lag features) and exogenous variables (external variables). Can be string identifier(s) or custom function(s).
​
float | None
default:"Equal weight among all data configs"
Relative sampling weight for this data source (normalized to sum to 1 across all data configs).
​
str | callable
default:"cross_entropy"
Loss function specification:
  • "cross_entropy": Chronos-style or text cross-entropy loss
  • "mse": Mean Squared Error (TimesFM-style)
  • "quantile" or "pinball": Quantile/Pinball loss (TiRex-style)
  • "multi_task": Multi-task learning (TimesFM 2.0-style)
  • Custom callable loss function
​
float
default:"0.1"
Portion of the dataset to use as validation data (0.0-1.0, where 1.0 means all data is validation).
At least one DataConfig must have validation_split < 1.0 for training to proceed.

​
OptimizationConfig

Configuration for training optimization parameters.
​
int
required
Total number of training steps for the experiment.
​
float
required
Maximum learning rate value.
​
int
required
Global batch size for training.
​
str | callable
default:"constant"
Learning rate scheduling strategy:
  • String options: "constant", "linear", "cosine", "exponential"
  • Custom function with signature: (learning_rate, current_step, total_steps) → decayed_rate
Warmup is applied after this schedule and must be disabled separately if not needed
​
int
default:"0"
Number of learning rate warmup steps.
​
int
default:"0"
Number of learning rate decay steps. Must be set to 0 when using custom learning rate schedules.
Constraint: warmup_steps + decay_steps ≤ total_training_steps
​
float | None
default:"max_learning_rate / 10"
Minimum learning rate value.
​
str
default:"Adam"
Optimizer algorithm. Options: "Adam", "SGD", "Lion"
​
float
default:"0.01"
L2 regularization coefficient.
​
float
default:"0.0"
Z-loss regularization coefficient. Set to 0.0 to disable.
​
float | None
default:"None"
Load balancing coefficient for Mixture of Experts (MoE) models. Only applicable for MoE architectures.
​
float
default:"1.0"
Gradient clipping threshold based on global L2 norm.

​
ModelConfig

Model architecture and initialization configuration.
​
str
required
Model architecture specification. Supports major dense and MoE Hugging Face architectures including Qwen, LLaMA, Gemma.
​
str
default:"normal"
Weight initialization strategy:
  • "none": Load from pre-trained model (Qwen/LLaMA/Gemma)
  • "normal": Normal distribution initialization
  • "xavier_uniform": Xavier uniform initialization
  • "wang_init": Wang initialization method
​
str | None
default:"None"
Path to pre-trained model for continual training. Must be None if init_method is not "none".
​
bool | None
default:"None"
Whether to load optimizer state from checkpoint. Set to True for continual training from checkpoint.
​
str
default:"fp16"
Model precision configuration:
  • "binary": Binary precision (1-bit)
  • "ternary": Ternary precision (1.58-bit)
  • "int2": 2-bit integer precision
  • "fp8": 8-bit floating point
  • "mxfp4": 4-bit microscaling floating point
  • "mxfp6": 6-bit microscaling floating point
  • "ue8m0": 8-bit unsigned integer with 0 exponent bits
  • "fp16": 16-bit floating point (default)
  • "fp32": 32-bit floating point