tw-niah-bench

Needle-in-a-haystack for Traditional Chinese, built from Taiwan.md.

420 cases across six context lengths (4k to 128k characters) and seven insertion depths (0% to 100%), ten needles per cell, drawn from 189 distinct source articles.

Part of OpenTWBench.

Why not the standard set

Standard NIAH uses English essays and an invented needle. Two things change here, and both matter.

The haystack is genuine Traditional Chinese prose. Long-context behaviour is not language-neutral: tokenisers split CJK very differently, and a model that retrieves reliably across 100k English tokens may not across the same number of Chinese characters. Almost nothing published measures this.

The needle is a real fact from a real article, not an invented sentence. An invented needle can only test retrieval. A real one also fails informatively: a model answering from its own knowledge of Taiwan rather than from the document is a different failure from one that cannot find the sentence.

Fields

Field	Description
question	The full prompt: the document, then the question
answer	The figure to return, with its unit
needle	The sentence inserted into the haystack
context_chars	Length of the assembled document
depth_percent	Where the needle sits, 0 = start, 100 = end
source_article	Article the needle came from
source_url	Link to that article
haystack_articles	How many articles the haystack was assembled from

Correctness is a construction invariant

No model is used to build this, so there is nothing generated to validate — only the construction, which is deterministic. Every case satisfies all of:

• the answer appears exactly once in the whole document, inside the needle
• the needle is present verbatim
• the answer does not appear in the question
• the needle contains exactly one number besides the year

That last rule took three attempts, and each earlier version passed the other checks while producing wrong items:

• 近一年均價約 96 萬一坪…2025 年 11 月有一筆成交來到每坪 105 萬 was keyed to 96 萬. The question asks for the figure given for 2025, which is 105 萬.
• 2026 年 5 月 2 日，編輯部以「30 年、30 人、30 個觀點」之名刊出引言頁 matched the number pattern only once, because 年 and 個 are not units — and offers a reader five digit runs to choose between.

Checking afterwards cannot catch either: the keyed answer still appears exactly once in the document, so the invariant holds on an item that is wrong. Counting every digit run at selection time makes the ambiguity impossible.

Licence

CC BY-SA 4.0, inherited from Taiwan.md, which is ShareAlike.

Source: Taiwan.md, CC BY-SA 4.0.