WikiText2DataLoader<T>

Models & Types · AiDotNet.Data.Text.Benchmarks

Loads the WikiText-2 language modeling dataset (≈ 2M tokens train, ≈ 245k val, ≈ 281k test from Wikipedia articles).

Inheritance: DataLoaderBase<T> → InputOutputDataLoaderBase<T, Tensor<T>, Tensor<T>> → WikiText2DataLoader<T>

How It Works

WikiText-2 expects:

Features are input token sequences Tensor[N, SequenceLength]. Labels are next-token target sequences Tensor[N, SequenceLength] (shifted by 1).

Uses the wikitext-2-raw-v1 archive flavor (preserves original casing and punctuation), which is the modern standard for subword-tokenizer pipelines. The companion WikiText103DataLoader uses the older tokenized flavor for backward compatibility — pick whichever matches your tokenizer.

Constructors

Constructor Summary
WikiText2DataLoader(WikiText2DataLoaderOptions options) Creates a new WikiText-2 data loader.

Properties

Property Summary
int FeatureCount { get; }
int OutputDimension { get; }
int TotalCount { get; }
string Description { get; }
string Name { get; }

Methods

Method Summary
Task<string> LoadRawTextAsync(DatasetSplit split, CancellationToken cancellationToken) Loads the raw, unprocessed text content for the requested split, auto-downloading via AutoDownload if the file is not already cached.
ValueTuple<IInputOutputDataLoader<T, Tensor<T>, Tensor<T>>, IInputOutputDataLoader<T, Tensor<T>, Tensor<T>>, IInputOutputDataLoader<T, Tensor<T>, Tensor<T>>> Split(double trainRatio, double validationRatio, int? seed)