Install libsvm with Homebrew, apt, dnf, MacPorts, Nix, pacman, zypper

Project links

• Homepage
• Upstream docs
• Package manager page

Commands and aliases

• svm-predict
• svm-scale
• svm-train

Project history

The project traces to 2000, the first year in the upstream copyright notice and the year identified by the 2011 ACM TIST paper as the start of active development. The official homepage describes LIBSVM as integrated software for support-vector classification, regression, and one-class distribution estimation, with multiclass classification, cross-validation, probability estimates, multiple kernels, and automatic model selection.

Version 2.8 incorporated an SMO-type algorithm using second-order working-set selection, tying the package to the JMLR work by Fan, Chen, and Lin. The project also kept a teaching posture: the upstream page promotes a practical SVM guide, an easy.py helper for beginners, the svm-toy visualization, and a standard sparse text data format that became familiar far beyond the original tarball.

Adoption history

LIBSVM spread through research and applied machine-learning workflows because it was easy to compile, easy to call from scripts, and easy to wrap. The upstream page lists Java, Python, R, MATLAB, Perl, Ruby, Weka, Common Lisp, Haskell, OCaml, LabVIEW, PHP, C#/.NET, and CUDA-related interfaces, plus inclusion in data-mining environments.

The package's adoption is also visible in distribution channels: the Homebrew formula is accompanied by Debian, Ubuntu, Fedora, MacPorts, Nix, Arch, and openSUSE package names in the input metadata, while upstream added PyPI installation support in the 3.25 release on April 14, 2021.

How it is used

The command-line workflow centers on svm-train, svm-predict, and svm-scale. Users put data in LIBSVM's sparse label/index/value format, scale features when needed, train a model, and run prediction; the README documents classification, regression, one-class SVM, kernels, class weights, cross-validation, and probability-estimate flags.

Library users link the C++ implementation or one of the language interfaces. Package users care that the upstream distribution remains small, permissively licensed, and stable enough that many higher-level tools can shell out to it or embed its data format.

Why package nerds care

LIBSVM is one of the classic examples of a research codebase becoming infrastructure because it did a narrow job well. It gave package managers a small, portable SVM implementation with stable command names and a data format that became a lingua franca for benchmarks.

Its significance is not only the algorithm implementation; it is the surrounding convention set: svm-train, svm-predict, svm-scale, heart_scale examples, grid search helpers, and a sparse text format that many datasets and wrappers learned to speak.

Timeline

• 2000: LIBSVM development begins, matching the start year in the upstream copyright notice and the 2011 ACM TIST project paper.
• 2005: Version 2.8 adopts an SMO-type algorithm based on second-order working-set selection.
• 2011: The ACM Transactions on Intelligent Systems and Technology article documents LIBSVM as a library for support vector machines.
• 2021: Version 3.25 adds PyPI installation support for the Python interface.
• 2025: Version 3.37 is announced on the upstream homepage on December 29, 2025.

Related projects

• LIBLINEAR is the closest sibling project from the same NTU group, aimed at large linear classification rather than kernel SVMs.
• scikit-learn, Weka, MATLAB/Octave wrappers, R interfaces, and many benchmark datasets relate to LIBSVM through bindings, compatible data formats, or workflow conventions.

Sources

• https://dl.acm.org/doi/10.1145/1961189.1961199
• https://github.com/cjlin1/libsvm
• https://raw.githubusercontent.com/cjlin1/libsvm/master/COPYRIGHT
• https://raw.githubusercontent.com/cjlin1/libsvm/master/README
• https://www.csie.ntu.edu.tw/~cjlin/libsvm/