CyrilB1531 

String distances, similarity metrics, tokenizers and text vectorizers for .NET — native, allocation-lean, no Python at runtime.

Sub-word tokenizers (WordPiece, SentencePiece, BPE and byte-level BPE), batch encoding, pooling and SIMD kNN search for .NET. No external dependency: ONNX inference lives in Lodestar.Onnx, so a caller who only... More information

The primitive types Lodestar's packages share: a compressed sparse row matrix with the vector and dense-block products a decomposition needs. No dependencies, no I/O, nothing to configure.

Classical hypothesis tests for .NET at scipy.stats parity: Student and Welch t, Mann-Whitney U, Wilcoxon signed-rank, chi-square, Fisher exact, Kolmogorov-Smirnov, one-way ANOVA, Kruskal-Wallis, Shapiro-Wilk,... More information

ONNX Runtime inference for Lodestar: runs a transformer encoder exported to ONNX and pools its token outputs into a sentence embedding. Split out of Lodestar.Embeddings so the tokenizers, the pooling and the... More information

Truncated SVD and non-negative matrix factorization for .NET at scikit-learn parity, over a sparse matrix and without centring it: latent semantic analysis on a term-document matrix, with explained variance,... More information

Clustering for .NET at scikit-learn parity: k-means by Lloyd's algorithm over a row-major span, with the initial centres accepted as an input so a run is reproducible entry by entry. Scores with... More information

String distances, similarity metrics, tokenizers and text vectorizers for .NET — native, allocation-lean, no Python at runtime.

Ordinary, weighted and generalized least squares with inference for .NET, at statsmodels parity: standard errors, t statistics, p-values, confidence intervals, R-squared and its adjusted form, the overall F... More information

Sub-word tokenizers, pooling and semantic (kNN) search for .NET. ONNX Runtime is isolated to this package so DataNet.Text stays dependency-free.

Applied fuzzy string matching for .NET (rapidfuzz-compatible fuzz ratios, process.extract, dedup). No external dependencies.

Classification metrics for .NET at exact scikit-learn parity: precision, recall, F1, FBeta, confusion matrix, classification report and ROC-AUC, with sklearn's averaging modes and zero-division semantics. No... More information

Applied fuzzy string matching for .NET (rapidfuzz-compatible fuzz ratios, process.extract, dedup). No external dependencies.

GPU kernels for Lodestar over ILGPU: a device-resident embedding matrix swept by a batch of queries, with cosine similarity and top-k computed on the accelerator. Additive — no Lodestar package depends on this... More information

Microsoft.Extensions.AI interoperability for Lodestar: exposes the ONNX embedding path as an IEmbeddingGenerator<string, Embedding<float>>, so a Microsoft.Extensions.AI or Semantic Kernel consumer can take... More information

Split conformal prediction for .NET at MAPIE parity: turn a point prediction into an interval, or a class into a prediction set, with a finite-sample coverage guarantee. Post-hoc over scores and labels — no... More information

Feature preprocessing for .NET at scikit-learn parity: fit a scaler on arrays and transform arrays, with the fitted statistics readable. Spans in, arrays out — no IDataView, no pipeline object, and nothing to... More information

Survival analysis for .NET, at lifelines parity: the Kaplan-Meier survival function with Greenwood variance and log-log confidence intervals, the Nelson-Aalen cumulative hazard, and the two-sample log-rank... More information

Classification metrics for .NET at exact scikit-learn parity: precision, recall, F1, FBeta, confusion matrix, classification report and ROC-AUC, with sklearn's averaging modes and zero-division semantics. No... More information

Math.NET Numerics interoperability for Lodestar: converts a CsrMatrix to and from Math.NET's sparse matrix in one pass over the stored values, so a caller already holding Math.NET types can reach... More information