Step Catalog

Every pipeline step that cursus supports — 61 steps — generated directly from the .step.yaml interface files. Each row links to that step’s page: its purpose, its inputs (with the upstream steps that can produce them), its outputs, and the downstream steps that consume it.

A cursus pipeline is a DAG of these steps. An edge is valid when a downstream step’s input type matches an upstream step’s output, and the upstream step is listed among the input’s compatible producers — see The DAG + Config → Pipeline model and Registry and Step Catalog.

The Compute column names how a step’s container is built: an SDK-managed DLC (sklearn / xgboost / framework / estimator / model), the SAIS script image, or byo_container — a user-supplied ECR image run verbatim (no image_uris.retrieve), which is how a non-DLC framework (e.g. GraphStorm/DGL) enters cursus. A step may also declare a per-step VPC (network_mode: config) to reach a VPC-only data source — shown on its page’s Compute section.

CradleDataLoading

Cradle data loading — pull source data via the SAIS/Cradle SDK.

Step	Node	Compute	Purpose	Consumes	Produces
CradleDataLoading	source	—	Cradle data loading step.	—	processing_output

RedshiftDataLoading

Redshift data loading — load data from Redshift via the SAIS SDK.

Step	Node	Compute	Purpose	Consumes	Produces
RedshiftDataLoading	source	—	Redshift SQL data loading step (source node with optional EDX upload).	—	processing_output

Processing

Processing jobs — data prep, feature engineering, evaluation, packaging.

Step	Node	Compute	Purpose	Consumes	Produces
ActiveSampleSelection	internal	sklearn	Active sample selection step that intelligently selects high-value samples from model predictions for Semi-Supervised Learning (SSL) or Active Learning workflows using confidence-based, uncertainty-based, diversity-based, or hybrid strategies.	processing_output	processing_output
BedrockBatchProcessing	internal	framework	Bedrock batch processing step that provides AWS Bedrock batch inference capabilities with automatic fallback to real-time processing for cost-efficient large dataset processing.	processing_output	processing_output
BedrockProcessing	internal	framework	Bedrock processing step that processes input data through AWS Bedrock models using generated prompt templates and validation schemas.	processing_output	processing_output
BedrockPromptTemplateGeneration	internal	sklearn	Bedrock prompt template generation step that assembles the prompt-config bundle (system prompt, category rules, output schema) into ONE standardized prompts.json prompt ruleset ({ruleset, rules} shape) with the output schema embedded in the prompt — the same contract the knowledge-routing producer emits, consumed by the Bedrock processing steps.	processing_output	processing_output
CurrencyConversion	internal	sklearn	Currency conversion processing step.	processing_output	processing_output
DataUploading	sink	—	Upload processed data to BDT (EDX/Andes) via SAIS SDK delegation.	processing_output	—
DummyDataLoading	internal	sklearn	Dummy data loading step that processes user-provided data instead of calling Cradle services.	processing_output	processing_output
DummyTraining	internal	framework	Training step that uses a pretrained model.	hyperparameters, model_artifacts	model_artifacts
EdxUploading	sink	script	Upload S3 data to EDX via EdxDataLoader (SINK node, no Kale required).	processing_output	—
FeatureSelection	internal	sklearn	Feature selection step using multiple statistical and ML-based methods with ensemble combination strategies.	processing_output	processing_output
GraphConstruction	internal	byo_container	GraphStorm gconstruct — build a partitioned DGL heterograph from the node/edge parquets + gconstruct schema emitted by GraphFeatureProcessing.	processing_output	processing_output
GraphFeatureProcessing	internal	byo_container	Turn per-seed subgraph pickles + labelled seeds into the GraphStorm GConstruct input (per-type node/edge parquets, reverse edges, node-ID-keyed masks, gconstruct_config.json).	processing_output	processing_output
GraphStormGNNInferenceEval	internal	byo_container	GraphStorm/DGL GNN out-of-time inference + evaluation on a custom GraphStorm GPU container (online-inference simulator → ROC-AUC / PR-AUC / Recall@Precision report + plots).	model_artifacts, processing_output	processing_output
GraphSubgraphExtraction	source	byo_container	Point-in-time k-hop subgraph pull from a property-graph DB for seed order IDs (BYO GraphStorm image, VPC-bound).	processing_output	processing_output
LabelRulesetExecution	internal	sklearn	Label ruleset execution step that applies validated rulesets to processed data to generate classification labels using priority-based rule evaluation with execution-time field validation.	processing_output	processing_output
LabelRulesetGeneration	internal	sklearn	Label ruleset generation step that validates and optimizes user-defined classification rules for transparent, maintainable rule-based label mapping in ML training pipelines.	processing_output	processing_output
LightGBMMTModelEval	internal	framework	LightGBM multi-task model evaluation step.	model_artifacts, processing_output	processing_output
LightGBMMTModelInference	internal	framework	LightGBM multi-task model inference step for prediction generation without metrics.	model_artifacts, processing_output	processing_output
LightGBMModelEval	internal	framework	LightGBM model evaluation step.	model_artifacts, processing_output	processing_output
LightGBMModelInference	internal	framework	LightGBM model inference step for prediction generation without metrics.	model_artifacts, processing_output	processing_output
MissingValueImputation	internal	sklearn	Missing value imputation step using statistical methods (mean, median, mode, constant) with pandas-safe values.	processing_output	processing_output
ModelCalibration	internal	sklearn	Calibrates model prediction scores to accurate probabilities.	processing_output	processing_output
ModelMetricsComputation	internal	sklearn	Model metrics computation step for comprehensive performance evaluation.	processing_output	processing_output
ModelWikiGenerator	internal	sklearn	Model wiki generator step for automated documentation creation.	processing_output	processing_output
Package	internal	sklearn	Model packaging step.	custom_property, model_artifacts, processing_output	model_artifacts
Payload	internal	sklearn	Payload testing step.	model_artifacts, processing_output	processing_output
PercentileModelCalibration	internal	framework	Creates percentile mapping from model scores using ROC curve analysis for consistent risk interpretation.	processing_output	processing_output
PiperMetricGeneration	internal	sklearn	PIPER metric generation step; recomputes ROC/PR curves and emits PIPER .metric + paired data CSVs flat to the output root for PIPER rendering.	processing_output	processing_output
PseudoLabelMerge	internal	sklearn	Pseudo label merge step that intelligently combines labeled base data with pseudo-labeled or augmented samples for Semi-Supervised Learning (SSL) and Active Learning workflows with split-aware merge, auto-inferred split ratios, and provenance tracking.	processing_output	processing_output
PyTorchModelEval	internal	framework	PyTorch model evaluation step.	model_artifacts, processing_output	processing_output
PyTorchModelInference	internal	framework	PyTorch model inference step for prediction generation without metrics.	model_artifacts, processing_output	processing_output
RiskTableMapping	internal	framework	Risk table mapping step for categorical features.	hyperparameters, processing_output	processing_output
SlipboxKnowledgeRouting	internal	framework	Slipbox knowledge routing step that hosts the DKS knowledge+ruleset corpus and runs compile→index→route internally, emitting a compiled prompt ruleset plus per-record routed rule names and routing confidence for downstream Bedrock processing.	custom_property, model_artifacts, processing_output	processing_output
StratifiedSampling	internal	sklearn	Stratified sampling step with multiple allocation strategies for class imbalance, causal analysis, and variance optimization.	processing_output	processing_output
TSAModelCalibration	internal	sklearn	TSA (Temporal Self-Attention) model calibration step using monotone B-spline calibration for converting raw prediction scores to well-calibrated probabilities for fraud detection.	processing_output	processing_output
TSAModelEval	internal	framework	TSA (Temporal Self-Attention) model evaluation step for dual-task PyTorch models with comprehensive metrics and visualizations.	model_artifacts, processing_output	processing_output
TSAPreprocessing	internal	framework	TSA (Temporal Self-Attention) data preprocessing step that performs sequence processing with feature transformation and scaling for fraud detection models.	processing_output	processing_output
TSATabularPreprocessing	internal	framework	TSA (Temporal Self-Attention) tabular preprocessing with explicit output declarations for processed_data and preprocessor artifacts.	processing_output	processing_output
TabularPreprocessing	internal	sklearn	Tabular data preprocessing step.	processing_output	processing_output
TemporalFeatureEngineering	internal	sklearn	Temporal feature engineering step that extracts comprehensive temporal features from normalized sequences for machine learning models.	processing_output	processing_output
TemporalSequenceNormalization	internal	sklearn	Temporal sequence normalization step for machine learning models with configurable sequence operations.	processing_output	processing_output
TemporalSplitPreprocessing	internal	sklearn	Temporal split preprocessing step with customer-level splitting and OOT validation.	processing_output	processing_output, training_data
TokenizerTraining	internal	framework	BPE tokenizer training step for customer name data with automatic vocabulary size tuning.	processing_output	processing_output
XGBoostModelEval	internal	xgboost	XGBoost model evaluation step.	model_artifacts, processing_output	processing_output
XGBoostModelInference	internal	xgboost	XGBoost model inference step for prediction generation without metrics.	model_artifacts, processing_output	processing_output
XgboostMtModelEval	internal	framework	XGBoost multi-task model evaluation step.	model_artifacts, processing_output	processing_output

Training

Training jobs — fit a model from prepared data + hyperparameters.

Step	Node	Compute	Purpose	Consumes	Produces
GraphStormGNNTraining	internal	byo_container	GraphStorm/DGL R-GCN GNN training on a partitioned heterograph, run in a bring-your-own GraphStorm ECR container.	hyperparameters, processing_output	model_artifacts, processing_output
LightGBMMTTraining	internal	estimator	LightGBM multi-task training with adaptive weighting and knowledge distillation.	hyperparameters, processing_output, training_data	model_artifacts, processing_output
LightGBMTraining	internal	estimator	LightGBM model training step using built-in algorithm.	hyperparameters, processing_output, training_data	model_artifacts, processing_output
PyTorchTraining	internal	estimator	PyTorch model training step.	hyperparameters, processing_output, training_data	model_artifacts, processing_output
SOPAInstructionTuning	internal	estimator	SOPA Stage 2 instruction fine-tuning step for BLIP2-based model (Q-Former + Phi-3 LLM) with tabular-to-text instruction following.	model_artifacts, training_data	model_artifacts, processing_output
TSATraining	internal	estimator	TSA (Temporal Self-Attention) model training step for PyTorch-based temporal attention models.	hyperparameters, training_data	model_artifacts, processing_output
XGBoostTraining	internal	estimator	XGBoost model training step.	hyperparameters, processing_output, training_data	model_artifacts, processing_output
XgboostMtTraining	internal	estimator	XGBoost multi-task training with one_output_per_tree strategy.	hyperparameters, processing_output, training_data	model_artifacts, processing_output

Transform

Batch transform jobs — run inference over a dataset with a model.

Step	Node	Compute	Purpose	Consumes	Produces
BatchTransform	internal	transformer	Batch transform step.	custom_property, processing_output	custom_property

CreateModel

Model creation — wrap trained artifacts into a deployable SageMaker model.

Step	Node	Compute	Purpose	Consumes	Produces
PyTorchModel	internal	model	PyTorch model creation step.	model_artifacts	custom_property
XGBoostModel	internal	model	XGBoost model creation step.	model_artifacts	custom_property

MimsModelRegistrationProcessing

Model registration — register a model with MIMS.

Step	Node	Compute	Purpose	Consumes	Produces
Registration	sink	—	Model registration step.	model_artifacts, payload_samples	—

Tuning

Hyperparameter tuning jobs — search a training step’s hyperparameters for the best model.

Step	Node	Compute	Purpose	Consumes	Produces
GraphStormGNNTuning	internal	byo_container	GraphStorm/DGL R-GCN hyperparameter tuning — a HyperparameterTuner search over the GNN training estimator, run in a bring-your-own GraphStorm ECR container.	hyperparameters, processing_output	model_artifacts

---

This catalog is generated from src/cursus/steps/interfaces/*.step.yaml by docs/gen_step_catalog.py. To change a step’s catalog entry, edit its .step.yaml and re-run the generator.