Eagle<T>

Models & Types · AiDotNet.VisionLanguage.InstructionTuned

Eagle: NVIDIA data-centric VLM with high-quality training data.

Inheritance: NeuralNetworkBase<T> → VisionLanguageModelBase<T> → Eagle<T>

Implements: IInstructionTunedVLM<T>

For Beginners

Eagle from NVIDIA focuses on getting the training data right rather than inventing complex new architectures. It systematically explores which design choices matter most for vision-language models — including vision encoder selection, projection strategies, and training data mixtures. The result is a model that achieves strong performance through careful engineering and high-quality data curation rather than architectural complexity. It uses a mixture of vision encoders to capture different aspects of visual information. Default values follow the original paper settings.

How It Works

Eagle (NVIDIA, 2024) takes a data-centric approach to vision-language model design, systematically exploring the design space of multimodal LLMs with a focus on high-quality training data curation. It uses a mixture of vision encoders to capture complementary visual features and demonstrates that careful data selection and training strategies matter more than architectural novelty for achieving strong multimodal performance.

References:

  • Paper: “Eagle: Exploring The Design Space for Multimodal LLMs” (NVIDIA, 2024)

Example

// Create an Eagle model for data-centric multimodal understanding
// using mixture of vision encoders with high-quality training
var architecture = new NeuralNetworkArchitecture<double>(
    inputType: InputType.TwoDimensional,
    taskType: NeuralNetworkTaskType.ImageClassification,
    inputHeight: 224, inputWidth: 224, inputDepth: 3, outputSize: 512);
            
// ONNX inference mode with pre-trained model
var model = new Eagle<double>(architecture, "eagle.onnx");
            
// Training mode with native layers
var trainModel = new Eagle<double>(architecture, new EagleOptions());

Constructors

Constructor Summary
Eagle(NeuralNetworkArchitecture<T> architecture, EagleOptions options, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> optimizer)
Eagle(NeuralNetworkArchitecture<T> architecture, string modelPath, EagleOptions options)

Properties

Property Summary
int DecoderEmbeddingDim { get; }
int EmbeddingDimension { get; }
int MaxGenerationLength { get; }
string LanguageModelName { get; }

Methods

Method Summary
ModelMetadata<T> GetModelMetadata()
Tensor<T> Chat(Tensor<T> image, IEnumerable<ValueTuple<string, string>> conversationHistory, string userMessage)
Tensor<T> EncodeImage(Tensor<T> image)
Tensor<T> GenerateFromImage(Tensor<T> image, string prompt) Generates text using Eagle’s data-centric multi-encoder fusion.