raz-systemone-nli-base (v7)

184M-param sibling of raz-systemone-nli-xsmall: same 725 NLI pairs from 110 labeled support-ticket states (repeat 8x) + 200,000 MNLI rows, 2 epochs, from cross-encoder/nli-deberta-v3-base. Powers the nli scorer in raz via --nli-model <dir>. 738MB.

Eval (all on states never seen in training)

split	judgments	xsmall (v5)	base (v7)
holdout C	60	.900	.983
holdout D	30	.833	.867
holdout E	30	.900	.833
holdout F	28	.893	.857
holdout G	29	.897	.897
MNLI (disjoint)	10000	.941	.980

Split decision: base nearly perfects the broad 60-state split (1 miss) but loses the small targeted splits to xsmall. Pick base for broad queries, xsmall for the best balance (and 10-min retrains vs ~2h).

Use

Same as xsmall: AutoModelForSequenceClassification, label 1 = entailment, premise + hypothesis per answer. Or in raz: hf download RazvanManolache/raz-systemone-nli-base --local-dir nli-base then --scorer nli --nli-model nli-base.

Limits

Same narrow-domain caveats as xsmall: 190 hand-written English support tickets; frustration tone is the weakest axis.