Models and techniques

Choose a model below, then use its parameter guide before changing an experiment. Each reference explains the supported native configuration, required inputs, output shapes, parameter effects, fitting and prediction lifecycle, reproducibility, resource limits and worked examples.

K nearest neighbors, K means, decision trees and contextual linear regression belong to aner.ml. Neural topology, differentiable operations and explicit optimizers belong to aner.nn; Dense classifier plans and training configurations use aner.ml. These are functions and native value types, not separately importable model submodules.

Choose a model

Model or technique	What it produces	Detailed reference	Runnable source
Natural language processing	Sparse lexical features, source linked evidence and exact neural token batches	NLP parameters and behavior	Classification, retrieval and neural workflows
K nearest neighbor classification	Class IDs, uniform vote fractions and nearest row inspection	KNN parameters and behavior	Complete KNN inspection
K means clustering	Centers, fitted assignments, new row cluster IDs and convergence readouts	K means parameters and behavior	Complete K means inspection
Decision tree classification	Class IDs, leaf vote probabilities and an inspectable numeric binary tree	Decision tree parameters and criteria	Decision tree workflow
Contextual linear regression	Numerical targets with fitted feature transformations, row identity and declared target meaning	FeatureLedger and linear regression	Building energy workflow
Dense neural classification	Trained Dense weights, fitted preprocessing, probabilities and classification reports	Dense model and training parameters	Training, validation and checkpoint continuation
POS and named entity tagging	Source aligned sequence labels, strict entities, error metrics and exact resumable training	Tagging parameters	POS and NER case studies
SGD and Adam	Explicit parameter updates and optimizer state	Optimizer parameters and state	Adam checkpoint continuation
Convolutional networks	NCHW image classifiers and configured architecture graphs	CNN layer, architecture and fitting parameters	CNN builders
Recurrent networks	RNN, LSTM and GRU sequence models, explicit state and supported advanced profiles	Recurrent model and fitting parameters	Recurrent workflow
Causal decoder and scientific generators	Native language model and bounded scientific generation experiments	Generative model and training parameters	Decoder training

The complete model API parameter index lists the supported signature families and explains each argument. Use machine learning, neural networks, sequence primitives, evaluation metrics and dataset partitions for the shared contracts.

Read a parameter reference

Required means the argument must be supplied to that native signature. A documented default exists only where an implemented overload supplies it. Examples are experiment settings, not hidden defaults or recommended values for every dataset.
Type and shape describe the value the checker or runtime accepts. A feature matrix, label column, logit matrix and cluster assignment can all be Tensor values while having different meanings and shapes.
Range and effect explain admissible values and how the parameter changes the algorithm. A parameter absent from a model's signature is not supported merely because another model has it.
Fitted state describes what the model retains and what the caller must keep separately. KNN's standardizer remains caller owned; a fitted Dense classifier owns its preprocessing.
Randomness identifies which operation owns the seed and whether repeated calls restart or continue it. KNN has deterministic row order ties and no seed parameter. Dense dropout has its own optional replay seed and persistent update schedule.
Checkpoint behavior is model specific. Dense, CNN, recurrent and generative formats have distinct contracts. Native KNN and K means do not gain serialization from the Dense checkpoint API.

Keep the comparison meaningful

Split training, validation and test roles before fitting preprocessing. Use validation to compare model settings; keep the final test partition out of tuning. Preserve original row mappings when extracting raw Tensor inputs, because those tensors no longer carry the Dataset role audit automatically.

Record the data source and row order, transformations, complete model settings, seeds and schedules, numerical backend, precision, runtime identity and outputs. A seed controls a documented random protocol; it does not make arbitrary software or device changes numerically identical. Compare probability losses with the intended label mapping and distinguish online training observations from detached evaluation metrics.

Availability of these references

These pages describe the reviewed implementation. Optional accelerator or distributed support depends on the installed build and available hardware. Dense dropout, its optional seed and version 7 checkpoint trace require the matching updated runtime. Updating documentation or an editor package does not publish a runtime or establish new binary availability. Check the runtime status and model specific capability notes when choosing an executable.