KEEP-Bench

KEEP-Bench (Key EvidencE Preservation) is a diagnostic benchmark for evaluating whether context compressors preserve the evidence needed to answer a question. Instead of inferring compression quality from downstream task accuracy, KEEP-Bench annotates each case with the evidence propositions that a faithful compression must retain, and checks each proposition directly in the compressed context.

The benchmark contains 4,102 cases in three settings, each targeting an evidence structure that compressors tend to break:

• Parallel evidence: the question contains underspecified mentions that admit two valid interpretations. Each interpretation leads through its own reasoning chain to a different answer, and both chains must be preserved.
• Conditional evidence: one hop of the reasoning chain holds only under an explicit qualifier, such as a period, edition, survey or version. The qualifier must be preserved together with the claim it scopes.
• Imbalanced evidence: an open-ended question has three or four valid answer perspectives. Dominant perspectives are supported by several paragraphs, while each minority perspective is supported by a single sentence. All perspectives must be preserved.

Paper (coming soon) · Code · Website

Dataset structure

Setting	Config	Cases	Subsets	Avg. paragraphs	Avg. tokens
Parallel	parallel	900	400 two-hop, 300 three-hop, 200 four-hop	22.8	2,799
Conditional	conditional	1,364	604 two-hop, 493 three-hop, 267 four-hop	21.0	2,544
Imbalanced	imbalanced_science, imbalanced_policy	1,838	Science: 1,010 (856 with four perspectives, 154 with three); policy: 828 (all with four perspectives)	9.8	3,577

Average context length is measured with the GPT-5 tokenizer.

Each config has a single test split. KEEP-Bench is an evaluation benchmark; there is no training split.

Loading

Access is gated, so log in first (hf auth login) and accept the terms on this page.

from datasets import load_dataset

parallel = load_dataset("uclanlp/KEEP-Bench", "parallel", split="test")
conditional = load_dataset("uclanlp/KEEP-Bench", "conditional", split="test")
science = load_dataset("uclanlp/KEEP-Bench", "imbalanced_science", split="test")
policy = load_dataset("uclanlp/KEEP-Bench", "imbalanced_policy", split="test")

The input to a compressor is question and context. context is the exact string used in our experiments; paragraphs holds the same content as structured records.

Fields: parallel

Field	Type	Description
id	string	Case id, e.g. parallel_3hop1_0001.
hop_type	string	MuSiQue reasoning type: 2hop, 3hop1, 3hop2, 4hop1, 4hop2, 4hop3.
hops	int	Number of hops per chain (2, 3 or 4).
question	string	Question containing the underspecified mention.
answer_1, answer_2	string	Gold answer under each interpretation.
logical_chain_1, logical_chain_2	list[string]	Evidence propositions of each reasoning chain, one per hop.
flattened_labels	list[string]	yes/no per proposition of logical_chain_1 + logical_chain_2. yes marks the propositions that distinguish the two referents; they define Para-Rec.
ambiguous_entities	list[string]	Underspecified mentions in the question.
unambiguous_entities	list[list[string]]	For each ambiguous mention, its two explicit referents. Passed to the judge as disambiguation constraints.
paragraphs	list[struct]	idx (int), title (string), paragraph_text (string), is_supporting (string: chain_1, chain_2 or distractor).
context	string	Paragraphs rendered as Document [i](Title: ...)text, one per line.

Fields: conditional

Field	Type	Description
id	string	Case id, e.g. conditional_2hop_0001.
hop_type, hops	string, int	As in parallel.
question_group	string	Cases sharing the same question, usually paired with different qualifiers (536 groups have two cases).
question	string	Question.
answer	string	Gold answer, stated together with its qualifier.
key	string	The qualifier, e.g. In the 1990s.
logical_chain	list[string]	Evidence propositions, one per hop. Exactly one proposition contains key; it is the qualified proposition.
disambiguation_proposition	string	States that the qualified relation differs across qualifiers. It explains the scope of the qualifier to the judge and is not an evaluated proposition.
paragraphs	list[struct]	idx, title, paragraph_text, is_supporting (string: supporting, disambiguating or distractor). The disambiguating paragraph shows that the relation is not valid without the qualifier.
context	string	Rendered as in parallel.

Fields: imbalanced_science and imbalanced_policy

Field	Type	Description
id	string	Case id, e.g. science_1major_3minor_0001 (1 dominant, 3 minority perspectives).
domain	string	Scientific field (science) or policy scenario (policy).
num_dominant, num_minority	int	Number of dominant and minority perspectives.
question_group	string	Cases sharing the same question with different dominant/minority configurations.
question	string	Open-ended question.
perspectives	list[string]	Answer perspectives, i.e. the evidence propositions. Policy perspectives are prefixed with their jurisdiction.
perspective_roles	list[string]	dominant or minority for each perspective, aligned with perspectives.
paragraphs	list[struct]	idx (int), text (string), label (string).
context	string	Paragraphs rendered as Document [i]text, one per line.

Paragraph label values refer to positions in perspectives:

• major_<i>_supporting: a chunk supporting dominant perspective i.
• minor_<i>_supporting: the chunk hosting the single sentence that supports minority perspective i. The host is a retrieved chunk into which that sentence was inserted, so the rest of the chunk may support another perspective.
• unsupporting: a chunk that supports none of the perspectives (123 chunks in 95 science cases).

Evaluation

The evaluation protocol, metrics (All-Rec, Para-Rec and Cond-Rec) and judge prompts are described in the paper and released in the code repository.

Email addresses and phone numbers in the web excerpts have been replaced with [EMAIL] and [PHONE]; these strings are not part of any evidence proposition.

License and terms of use

KEEP-Bench is released under CC BY-NC 4.0.

• The parallel and conditional settings are derived from MuSiQue (CC BY 4.0).
• Questions, evidence propositions, annotations and rewritten passages were created by the KEEP-Bench authors.
• Passages in the imbalanced setting are excerpts from publicly accessible web pages. Copyright remains with the original authors. These excerpts are provided solely for non-commercial research on context compression and may not be redistributed separately.

If you are a rights holder and want content removed, please open a discussion in the Community tab of this repository, and we will remove it promptly.

Citation

@article{keepbench2026,
  title   = {{KEEP-Bench}: Evaluating Evidence Preservation in Context Compression},
  author  = {TBD},
  journal = {arXiv preprint arXiv:TBD},
  year    = {2026}
}