RecurrentGemmaOptions
Options & Configuration · AiDotNet.NeuralNetworks.Options
Configuration options for the RecurrentGemmaLanguageModel.
Inheritance: ModelOptions → NeuralNetworkOptions → ModelHyperparameterOptions → SequenceModelOptions → RecurrentGemmaOptions
How It Works
Every value here is a PAPER DEFAULT assigned in the constructor rather than baked into a property initializer, so a caller can replace any of them without having to reconstruct the rest. The defaults reproduce RecurrentGemma (Botev et al., 2024); the type carried no settings at all before, so the model silently inherited whatever the base supplied.
Constructors
| Constructor | Summary |
|---|---|
RecurrentGemmaOptions() |
Initializes an options instance with the paper’s defaults. |
RecurrentGemmaOptions(RecurrentGemmaOptions other) |
Initializes an options instance by copying another. |
Properties
| Property | Summary |
|---|---|
bool EnableGradientClipping { get; set; } |
Gets or sets whether global-norm gradient clipping is enabled. |
bool ScaleEmbeddingsBySqrtWidth { get; set; } |
Gets or sets whether the input embeddings are multiplied by the square root of the model width. |
double Beta1 { get; set; } |
Gets or sets AdamW’s first-moment decay. |
double Beta2 { get; set; } |
Gets or sets AdamW’s second-moment decay. |
double Epsilon { get; set; } |
Gets or sets AdamW’s numerical-stability epsilon. |
double LearningRate { get; set; } |
Gets or sets the AdamW learning rate. |
double MaxGradientNorm { get; set; } |
Gets or sets the maximum gradient norm. |
double WeightDecay { get; set; } |
Gets or sets the decoupled AdamW weight decay. |