YuvGev/hebrew-toolbench

Hebrew ToolBench

A Hebrew tool-calling benchmark, modelled on BFCL: 800 single-turn items in Hebrew, 52 Israeli-domain tools, and a deterministic AST scorer (no LLM judge).

It tests what generic benchmarks miss:

• Hebrew mentions that an API needs in English or as an enum ("באר שבע" → Beer Sheva);
• the Jewish calendar and gematria;
• ₪ amounts written in words;
• DD/MM dates and spoken times;
• gendered inflection;
• right-to-left text with embedded IDs.

Code, scorer and raw model outputs: github.com/YuvGev00/hebrew-toolbench

Quick start

import json
from datasets import load_dataset

ds = load_dataset("YuvGev/hebrew-toolbench", split="test")
item = ds[0]
functions = json.loads(item["functions"])  # tool schemas to offer the model
gold = json.loads(item["gold"])  # expected call(s), with every acceptable value
print(item["question"])

To score a model, use the AST scorer in the GitHub repo (uv run hebrew-toolbench run and report). It handles the acceptable-value lists, defaults and order-free parallel calls.

Fields

Field	Type	Meaning
id	string	Item id, e.g. simple_001
category	string	simple (300), multiple (200, pick 1 of 3 tools), parallel (150, 2–4 calls), irrelevance (150, no call is right)
tags	list of strings	What the item tests, e.g. canonical_value, hebrew_calendar, currency_ils, date_ddmm, time_words, mixed_direction
system	string	System prompt. Every item is anchored to Tuesday 20 Oct 2026, 09:00 (ט׳ בחשוון תשפ״ז), so relative dates have one answer
question	string	The user's request, in Hebrew
functions	JSON string	Tool schemas (OpenAI function format)
gold	JSON string	Expected calls. Each argument lists every acceptable value; [] means no call is correct
source	string	v0 (the first 150 items) or the domain pack the item came from
review	string	Native-speaker review status (currently pending for every item)

Results (v1, 2026-10-22)

Accuracy in %, with a 95% Wilson interval. Single turn, tool_choice: auto.

Model	Overall	95% CI	simple	multiple	parallel	irrelevance
claude-sonnet-5-5 †	95.8	94.1–96.9	96.3	98.5	95.3	91.3
gemini-3.5-flash	93.4	91.4–94.9	96.7	97.5	95.3	79.3
claude-haiku-5-5 †	92.5	90.5–94.1	93.7	95.0	90.0	89.3
gemma-4-31b-it	90.6	88.4–92.5	89.3	91.5	91.3	91.3
deepseek-v4.1-flash	90.1	87.9–92.0	92.7	91.0	90.7	83.3
glm-5.3	90.1	87.9–92.0	91.0	89.0	90.0	90.0
qwen3.8-27b	85.9	83.3–88.1	84.3	85.5	85.3	90.0
gemini-3.5-flash-lite	81.8	78.9–84.3	85.0	90.0	82.0	64.0
minimax-m3.1-flash:free	59.9	56.4–63.2	61.3	65.5	70.7	38.7

† The items were drafted by Claude (see below), so the two Claude models may have a home advantage.

Main findings:

• The Hebrew calendar is the hardest part for every model. Converting a Hebrew date such as י״ט בכסלו to a Gregorian date drops most models far below their plain simple score.
• Near-miss requests cause spurious calls. An offered tool that almost fits (98 octane with a tool that only knows 95) brings models down to 59–89%.
• Writing values in Hebrew where the API expects English appears mainly in the smallest model: 19 of Gemini Flash-Lite's 71 errors on canonical-value items.

How the items were made, and limitations

• Drafting: 650 of the 800 items were drafted by an LLM (Claude) under written authoring rules: native colloquial Hebrew, exactly one correct answer, computed golds. They are not translations.
• Automatic checks: 2,454 tests in CI. Every gold is valid against its schema, every Hebrew date exists (computed with pyluach), and questions are unique.
• Adjudication: every item that both strong models failed (19) was reviewed by hand; 8 were fixed.
• Native review is pending: no independent native speaker has signed off yet. Treat v1 as a preview.
• Single turn only: there are no multi-turn or executable checks.

License

Data: CC BY 4.0. Code (in the GitHub repo): MIT.

Citation

@misc{geva2026hebrewtoolbench,
  title  = {Hebrew ToolBench: a Hebrew tool-calling benchmark},
  author = {Geva, Yuval},
  year   = {2026},
  url    = {https://github.com/YuvGev00/hebrew-toolbench}
}