Nativ-Bench

Nine Hebrew test sets in one format: a state (any Hebrew text), a question and 2 to 7 options, one of them correct. Models are scored on accuracy and on calibration (expected calibration error, ECE). Any model that gives a probability for each option can be evaluated. It was made for Nativ.

Tasks

Public test sets, built from their sources by build_public.py (not redistributed here):

task	question	options	items	source	license
topic_sib200	topic of the sentence	7	204	SIB-200 heb_Hebr test	CC BY-SA 4.0
sentiment_nnlp	positive / negative / neutral	3	2,156	HebArabNlpProject/HebrewSentiment test, one item per text	see source
nli_hebnli	the second sentence follows / contradicts / can't tell	3	884	HebNLI gold test, Hebrew labels	see source
reading	which answer does the passage support	4	900	Belebele heb_Hebr	CC BY-SA 4.0

Included in data/:

task	question	options	items	test data	license
routing	which of 4 intent descriptions matches the request	4	2,973	MASSIVE he-IL test; Hebrew descriptions of the 60 intents, 3 random wrong intents per item	CC BY 4.0
sentiment	positive / negative / off-topic	3	1,695	OnlpLab Hebrew-Sentiment-Data, test_gold (token version)	MIT
grounded	does the paragraph contain the answer to the question	2	864	HeQ v1.1 test, answerable and unanswerable questions, balanced	CC BY 4.0
pii_natural	does the message contain personal information	2	656	messages written by DictaLM-3.0-24B in held-out scenarios (city hall, airline, hotel, vet ...), balanced	CC BY 4.0
policy	does the request meet the rule: meets / violates / missing information	3	1,000	4 rule types (refund, cancellation, discount, free delivery) with held-out thresholds; labels computed in code, balanced	CC BY 4.0

Notes:

• routing, sentiment and grounded use the test splits of their sources. The policy test uses thresholds that do not appear in Nativ's training data, and the pii_natural scenarios differ from the ones it trained on.
• pii_natural keeps only messages where the 24B's answer matches the label it wrote.
• MASSIVE he-IL and HebNLI are translated from English.

Format

One JSON object per line:

{"kind": "policy", "source": "synthetic-policy-test", "source_id": "policy-test-0",
 "state": "מדיניות: משלוח חינם בהזמנה מ-250 ש\"ח ומעלה.\nהבקשה: הזמנתי מכנסיים ב-89 ש\"ח וחולצה ב-183 ש\"ח. מגיע לי משלוח חינם?",
 "question": "האם הבקשה עומדת במדיניות?", "options": ["עומדת", "לא עומדת", "חסר מידע"], "answer_index": 0}

answer_index is 0-based. Labels are balanced in grounded, pii_natural, policy and nli_hebnli; the other tasks keep their sources' label shares (sentiment is 79% positive). Answer positions are shuffled.

Evaluation

pip install datasets && python build_public.py      # creates the 4 public test sets in data/
# predictions/<task>.jsonl, one line per item: {"source_id": ..., "probabilities": [p_option_1, p_option_2, ...]}
python evaluate.py predictions/

• The prediction is the option with the highest probability (the first one on a tie).
• ECE uses the probability of the predicted option, in the bins [0, .5), [.5, .6), [.6, .7), [.7, .8), [.8, .9), [.9, 1.0]: the average over bins, weighted by item count, of |accuracy - mean probability|.
• A model that outputs a single answer can be scored with probability 1 on that answer; its ECE is then 1 - accuracy.

Results

Accuracy, with ECE in parentheses; the best result in each row in bold.

task	Nativ v2	OpenAI Decisions	Laya multilingual	Laya-Hebrew	DeepSeek V4 Flash
topic_sib200	0.858 (.072)	0.858 (.058)	0.583 (.112)	0.789 (.096)	0.858 (.109)
sentiment_nnlp	0.656 (.105)	0.699 (.128)	0.429 (.146)	0.830 (.062)	0.611 (.282)
nli_hebnli	0.808 (.089)	0.734 (.062)	0.480 (.191)	0.606 (.114)	0.577 (.242)
reading	0.588 (.133)	0.910 (.016)	0.318 (.229)	0.758 (.026)	0.883 (.086)
routing	0.965 (.011)	0.925 (.011)	0.644 (.085)	0.887 (.028)	0.905 (.050)
grounded	0.855 (.019)	0.737 (.140)	0.582 (.338)	0.694 (.061)	0.847 (.096)
sentiment	0.905 (.009)	0.865 (.076)	0.680 (.062)	0.822 (.064)	0.838 (.132)
pii_natural	0.954 (.078)	0.593 (.285)	0.566 (.356)	0.848 (.028)	0.890 (.069)
policy	0.926 (.060)	0.970 (.038)	0.298 (.373)	0.406 (.365)	0.624 (.311)

Which of these datasets each model trained on: Nativ trained on the train splits of MASSIVE (routing), HeQ (grounded), OnlpLab (sentiment) and SIB-200 (66 sentences that are not part of Belebele), and on MultiNLI, which the HebNLI test was translated from (every premise in the HebNLI test left out); it was also trained on the pii_natural and policy task types, with other scenarios and other thresholds. Laya-Hebrew trained on HebNLI, HeQ and Hebrew sentiment data, among others. OpenAI Decisions (gpt-6-luna, public beta, API) and Laya multilingual have not published their training data. OpenAI Decisions declined a few requests (about 1% of routing); they count as a uniform guess. DeepSeek V4 Flash was scored zero-shot by the probability of each option's letter.

Changes

• 2026-10-08: nli_hebnli is now the whole HebNLI gold test (884 pairs) with the linguists' Hebrew labels (hebrew_label). It used to be a 579-item sample built from the English labels (original_label): 305 rows that spell the label with a capital letter were dropped, and 55 of the rest have a different Hebrew label.

License

The files in data/ are CC BY 4.0, except sentiment.jsonl (MIT, from OnlpLab Hebrew-Sentiment-Data). Sources: MASSIVE (FitzGerald et al., 2022, Amazon), HeQ (Cohen et al., 2023, NNLP-IL) and Hebrew-Sentiment-Data (Amram et al., 2018, OnlpLab). The public test sets keep their own licenses: SIB-200 (Adelani et al., 2024), HebrewSentiment and HebNLI (HebArabNlpProject), Belebele (Bandarkar et al., 2023, Meta). Generated items were written with DictaLM-3.0 models (Dicta, Apache-2.0).