NeedleBench

OpenCompass long-context suite that plants synthetic needles in English and Chinese haystacks at chosen lengths and depths, plus an Ancestral Trace Challenge.

Also known as: NeedleBench v1, needlebench

unassessed

This page is a discovery lead. Nobody has yet assessed it against the catalogue contract, so it carries no disposition. Absence of evidence here is not evidence of staleness.
Categorylong-context
Subcategorybilingual synthetic needle retrieval, multi-needle reasoning, and ancestral trace
Page statussuperseded
MetricOpenCompass v1: keyword hit 100 else 0.2 x Levenshtein similarity; v1 paper overall weights 0.4/0.3/0.3 on S-RT/M-RT/M-RS
Directionhigher_is_better
Unit%
Dataset licenceMIT
PublisherShanghai AI Laboratory and Tsinghua University

What it measures

NeedleBench builds a long prompt and asks a model to recover or reason over planted facts ("needles"). Information-sparse tasks put one or more synthetic facts into a haystack of Paul Graham essays (English) or ChineseDomainModelingEval passages (Chinese), at a chosen token length and depth. Single-needle retrieval (S-RT) asks for one fact. Multi-needle retrieval (M-RT) asks for several. Multi-needle reasoning (M-RS) asks the model to combine two to five facts. The Ancestral Trace Challenge (ATC) is information-dense: every sentence is a kinship fact and there is no filler. English and Chinese. Text only. OpenCompass v1 length packs are 4k, 8k, 32k, 128k, 200k and 1000k (the 256k pack is in needlebench_v2, not this tree).

Task format

Long generated prompt in; short free-text answer out. OpenCompass v1 sparse scoring (NeedleBenchOriginEvaluator) is 100 if a core keyword appears, else 0.2 times Levenshtein similarity to the reference. ATC in the later paper uses boxed exact match; v1 configs do not. OpenCompass v1 uses NeedleBenchOriginDataset, a GPT-4 tokenizer length target, and a length buffer.

Models reporting this benchmark

No model card in ModelSpec reports this benchmark yet.

Data

This page as JSON · Edit on GitHub