Relational Transformers

Documentation

  • Installation
    • Install with uv
    • Install with pip
    • Install with Conda
    • Install from Source
    • Editable Install
    • Install PyTorch with CUDA support
    • Model Files and Caching
  • Testing
    • Coverage
    • Opt-in Suites
    • Linting and Documentation Checks
    • Integration Testing
  • Quickstart
    • Load RT-J
    • Encoding Cells
    • Typed Cells
    • Batch Prediction
    • Next Steps
  • Examples
    • Prediction and analysis
    • Training
    • Deployment
  • Computing Predictions
  • Prediction
    • Activations
    • Predicting with a Fitted Task Head
    • Contextual Cell Embeddings
    • Model Outputs
  • Relational batches
    • Fields
    • How the Model Reads a Batch
    • Constructing a Batch
    • All-text Convenience Batches
    • Device Movement and Ablation
  • Backends
    • PyTorch
    • Triton
    • ONNX
    • Meta
  • Efficiency
    • Backend Selection
    • Batch Similar Context Lengths
    • Quantized Models
    • ONNX Export
    • Measure End-to-End Latency
  • Ablation
    • How Ablation Works
    • Choosing What to Ablate
    • Measuring Over a Dataset
  • Custom models
    • Checkpoint Resolution
    • Saving Checkpoints
    • Working with RTJModel Directly
    • Changing the Embedding Space
  • Pretrained Models
    • Published Models
    • Model Input Contract
    • Classification and Ranking
    • Regression and Forecasting
  • Dataset Overview
    • Accepted Input Types
    • Dataset Construction
    • Splitting Relational Data
    • Pre-existing Datasets
  • Loss Overview
    • Loss Table
    • Classification
    • Regression and Forecasting
    • Custom Loss Functions
  • Training Overview
    • Why Fine-tune?
    • Training Components
    • Model
    • Dataset
      • Dataset Format
    • Loss Function
    • Training Arguments
    • Evaluator
    • Trainer
    • End-to-End Example
  • Choosing a Training Path
    • Head Tuning
    • Full Fine-tuning
    • Recommendation
  • Task-head tuning
    • Tuning Knobs
    • Reloading a Saved Head
  • Full-model fine-tuning
    • What a Training Step Does
    • Triton-compiled GPU training
    • Practical Notes
  • Training Examples
    • Head Tuning
    • Multiclass Classification
    • Full-model Fine-tuning
    • Evaluation During Training
    • Custom PyTorch Loop
  • Model API
    • RelationalTransformer
      • RelationalTransformer
    • RTJModel
      • RTJModel
    • ModelOutput
      • ModelOutput
  • Batch API
    • RelationalBatch
      • RelationalBatch
  • Datasets
    • RelationalDataset
      • RelationalDataset
  • Evaluation
    • BinaryClassificationEvaluator
      • BinaryClassificationEvaluator
    • RegressionEvaluator
      • RegressionEvaluator
    • AblationEvaluator
      • AblationEvaluator
    • SequentialEvaluator
      • SequentialEvaluator
  • Losses
    • BinaryClassificationLoss
      • BinaryClassificationLoss
    • MulticlassClassificationLoss
      • MulticlassClassificationLoss
    • MultilabelClassificationLoss
      • MultilabelClassificationLoss
    • RegressionLoss
      • RegressionLoss
    • loss_for
      • loss_for()
  • Training API
    • RelationalExample
      • RelationalExample
    • RelationalTrainingArguments
      • RelationalTrainingArguments
    • RelationalTrainer
      • RelationalTrainer
    • TaskHead
      • TaskHead
    • fit_head
      • fit_head()
  • ONNX API
    • export_onnx
      • export_onnx()
    • OnnxBackend
      • OnnxBackend
Relational Transformers
  • Search

© Copyright 2026, RelativeDB.

Built with Sphinx using a theme provided by Read the Docs.