release: 0.2.0

Nothing here has ever been tagged, and the CHANGELOG had 78 unreleased
entries: everything from this modernization -- the correctness fixes, the
evaluation data, the domain tasks, the 28% speedup -- was reachable only
by cloning the repository. `pip install hyperhyper` still served the 2019
code, with the bugs.

0.2.0 rather than 0.1.2 because reported numbers move, and rather than
1.0 because the new API has no usage behind it yet. The CHANGELOG now
opens with why the numbers moved and says plainly not to compare a 0.1.x
score with a 0.2.0 one: analogy accuracy used to be structurally 0.0,
similarity gold was ranked as strings, the cache key ignored **kwargs so
results depended on call order, and subsample="prob" used word2vec's
discard probability as a keep probability. Same-version results stay
reproducible, and the tokenizer identity is recorded per result so old
and new numbers cannot collide silently.

Artifacts verified rather than assumed: the wheel installs into a clean
environment, reports 0.2.0, trains and evaluates, and carries all 23
bundled datasets across en/de/fr/sv/da; it contains no tests/, bench/ or
tools/. The sdist does carry those 24 entries, so it can run its own
tests -- which it could not before b597904.

Author: jfilter

CHANGELOG.md

@@ -2,8 +2,22 @@
 
 ## Unreleased
 
-Modernization of the package for current Python and dependency versions.
-Everything below is user-visible; several items change numeric results.
+## 0.2.0 - 2026-07-22
+
+Modernization of the package for current Python and dependency versions, plus
+the bug fixes and evaluation-data work that came out of it. Everything below is
+user-visible.
+
+> **Reported numbers move in this release, and that is the point.** Several of
+> the fixes below were wrong *answers*, not slow ones: word-analogy accuracy was
+> structurally 0.0, word-similarity gold scores were ranked as strings, the
+> cache key ignored `**kwargs` so results depended on call order, and
+> `subsample="prob"` used word2vec's *discard* probability as a *keep*
+> probability. The evaluation data was also cleaned and extended. **Do not
+> compare a score from 0.1.x with a score from 0.2.0** -- they are not measuring
+> the same thing. Scores computed with the same version remain reproducible, and
+> the tokenizer identity is now recorded with every result so old and new
+> numbers cannot silently collide.
 
 ### Added
 

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "hyperhyper"
-version = "0.1.1"
+version = "0.2.0"
 description = "Python package to construct word embeddings for small data"
 readme = "README.md"
 requires-python = ">=3.10,<3.14"