0.2.1 (2026-08-21)

• Fix the build against libxgboost >= 3.1: drop a stray, unused binding to XGDMatrixCreateFromCSREx, which was removed from the C API after 3.0 and broke ctypes stub compilation on newer libxgboost (e.g. Homebrew's 3.4.1). The sparse-CSR path already uses the modern XGDMatrixCreateFromCSR.

0.2.0 (2026-08-20)

First opam release.

• Native dependency is now hybrid: the build links a system libxgboost ≥ 3.0 when pkg-config finds one (installed via opam depexts on platforms that package it), and otherwise compiles a pinned, vendored copy of XGBoost from source (bundled under vendor/xgboost, CMake, static link, no network required). This lets opam install xgboost work on any platform, including those without a distro package.
• Xgboost.Eval — parse eval_one_iter metric strings; compute AUC and ROC directly from prediction/label Bigarrays.
• Xgboost.Cv.k_fold — k-fold cross validation with optional group-coherent splitting (?group_ids) and summarise.
• Xgboost.DMatrix.slice — row-subset a DMatrix.

0.1.0

Initial release. Native OCaml bindings to libxgboost ≥ 3.0 covering:

• DMatrix construction

  • dense (Bigarray.Array2 of float32, zero-copy via the modern XGDMatrixCreateFromDense array_interface path)
  • sparse CSR (modern XGDMatrixCreateFromCSR)
  • streaming iterator (DMatrix.of_iterator) backed by XGProxyDMatrixCreate + XGDMatrixCreateFromCallback, with optional external-memory caching via cache_prefix

• Booster lifecycle, training, prediction

  • update_one_iter, boost_one_iter (custom gradient/hessian), eval_one_iter, reset
  • predict (eager copy), predict_dense (in-place from Bigarray), Unsafe.predict_borrowed (no-copy expert variant)
  • feature_score with selectable importance type
• Persistence: model save/load via path or buffer; JSON config save/load
• GC-safe lifetime model: Gc.finalise_lastd handles, scoped combinators (with_), idempotent free
• Comprehensive test suite (alcotest + qcheck), cross-layer fixture parity oracle, clean run under AddressSanitizer
• Benchmarks against pure-C reference and Python xgboost on a fixed workload grid; see BENCH.md