v0.26

Full Changelog: v0.25...v0.26

v0.25

Full Changelog: v0.24...v0.25

v0.24

Full Changelog: v0.23...v0.24

v0.23

Full Changelog: v0.22...v0.23

v0.22

Full Changelog: v0.21...v0.22

v0.21

Full Changelog: v0.20...v0.21

v0.20

Changes since v0.19

• Emoji tokens now use literal characters (emoji:💰) instead of hex codepoints for readability
• Rule features use log(count+1) for tflags multiple rules instead of binary, preserving hit count signal
• Merge attach: prefix into body: to reduce vocabulary duplication

v0.19

Changes since v0.18

• Add emoji tokenization: extract individual emojis as emoji:XXXX tokens using Unicode codepoints
• Merge attach: prefix into body: for text tokenization, reducing vocabulary duplication while keeping subj:, link:, image:, and from: prefixes separate

Training improvements (in sa-fann-train)

• Add --hidden2/--hidden3 for multi-layer networks
• Add --prune-method df with --min-df/--max-df for document-frequency-based vocabulary and rule pruning
• Add --log for detailed pruning output
• --skip-rbl-checks now also disables URIBL, DCC, Pyzor, and Razor for consistent network-free training

New utilities

• bin/add-mxg-flags — query sender reputation API and update X-MXG-Flags headers in corpus messages
• bin/strip-mxg-flags — remove specified flags from X-MXG-Flags headers
• bin/fann-experiment — hyperparameter grid search with parallel workers

v0.18

Full Changelog: v0.17...v0.18

v0.17

Full Changelog: v0.16...v0.17