Skip to content

DKSplit v1.0.0

Choose a tag to compare

View all tags
@ABTdomain ABTdomain released this 10 Jun 09:26
v1.0.0
df4fb98

DKSplit v1.0.0 — First stable release

The model is frozen and the public API is stable: split, split_batch, split_topk / split3 / split5.

Added

  • Top-k API: split3() / split5() / split_topk(text, k) return the k best
    candidate segmentations, ranked best-first. Decoded with k-best Viterbi over the
    same CRF - no model change, no new dependencies.

Changed

  • Benchmark restructured: removed digit-driven and pure-noise samples;
    added a might_right field for genuinely ambiguous cases (Lenient EM).
    Numbers are not comparable to previously reported results.
    • Strict EM: 86.5% · Lenient EM: 91.5% (1,000 hand-audited samples)
    • Top-3 coverage: 98.5% · Top-5 coverage: 99.3%
  • Dataset and evaluation script: /benchmark

Stability guarantees

  • Model weights are frozen - identical input → identical output across patch releases.
  • Public API will not break within v1.x (SemVer).

PyPI: https://pypi.org/project/dksplit/1.0.0/

Assets 2
Loading