groundedness / eval /groundedness_eval.md
bogdanraduta's picture
Upload eval/groundedness_eval.md with huggingface_hub
9e665d5 verified
|
Raw
History Blame Contribute Delete
2.18 kB

groundedness evaluation

Threshold 0.5, taken from the config.

Pair accuracy: 0.899 (both halves of a minimal pair correct. This is the headline number for this model.)

Single-label head, so this table reports exact-match accuracy. Precision, recall and F1 over did the detector fire are meaningless here: argmax always names a class, so every row would count as a true positive and every language would read 1.000 whether or not a single prediction was right.

Per language

Language N Accuracy Notes
bg Bulgarian 80 0.975
cs Czech 78 0.962
da Danish 80 0.963
de German 78 0.936
el Greek 80 0.950
en English 78 0.897
es Spanish 80 0.950
et Estonian 76 0.961
fi Finnish 78 0.936
fr French 80 0.938
ga Irish 78 0.974
hr Croatian 78 0.923
hu Hungarian 80 0.950
it Italian 80 0.988
lt Lithuanian 80 0.963
lv Latvian 80 0.950
mt Maltese 80 0.938 not in base model pretraining
nl Dutch 80 0.938
pl Polish 80 0.912
pt Portuguese 80 0.963
ro Romanian 80 1.000
sk Slovak 80 0.925
sl Slovenian 80 0.963
sv Swedish 80 0.912
tr Turkish 78 0.962
az Azerbaijani 80 0.912

The base-model note is a fact about pretraining, not a cause of the score beside it. nsfw Maltese carried the same note at 0.000 and reached 1.000 on corpus size alone, with nothing about the base model changed. Check how many examples a weak score rests on before reaching for this.

Per register

Register N Accuracy
lexical_overlap 412 0.949
negation_conflict 414 0.947
numeric_conflict 404 0.908
scope_conflict 416 0.966
unstated 416 0.966

Known weaknesses

The three weakest languages by accuracy: en at 0.897, az at 0.912, pl at 0.912.

These are published rather than dropped. A coverage table with the bad rows removed is not a coverage table.