-
[NA] [SDK] fix: forward char_order and ignore_whitespace to NLTK in ChrF metric (#7230)
frostbyte_neo 发布于
2026-06-25 08:54:31 +00:00 - fix: forward char_order and ignore_whitespace to NLTK in ChrF metric
The ChrF metric stored char_order and ignore_whitespace but only passed beta
to nltk.translate.chrf_score.sentence_chrf, so those parameters were silently
ignored (and ignore_whitespace's documented default of False was overridden by
NLTK's own default of True). Forward char_order as max_len and ignore_whitespace
to NLTK. Clarify in the docstring that word_order (chrF++) is not supported by
the NLTK backend and requires a custom chrf_fn. Add a regression test asserting
the parameters are forwarded.- test: exercise ChrF config through the public API and follow naming convention
Address review: rename the test to the test_WHAT__CASE__EXPECTED pattern and
assert on public ChrF.score output (scores change when char_order /
ignore_whitespace vary) instead of inspecting the internal NLTK backend call
args. Gated with importorskip('nltk') since the NLP backend is optional and not
a test dependency.- fix: default ChrF ignore_whitespace to True to preserve prior behaviour
Address review: before forwarding the flag, NLTK's own default (True) always
applied, so whitespace was effectively ignored. Defaulting to True keeps results
stable for users who relied on that implicit behaviour across an Opik update.下载附件