cabbage972/corpus-assay-benchmark-registry

corpus-assay benchmark registry

25 pinned specs of LLM evaluation benchmarks, for benchmark decontamination of LLM pretraining data. Each row says which Hugging Face dataset, revision, config, split and text fields to index so that a decontamination scan protects the items an evaluation harness actually scores.

This is the built-in registry of corpus-assay 0.2.0, a Rust-accelerated tool that finds benchmark text in LLM training corpora (n-gram overlap, with per-benchmark and per-item attribution) and writes cleaned shards. The rows hold no benchmark text, only pointers to it.

Why a registry

Getting the spec wrong silently weakens decontamination:

• Wrong split. HellaSwag, BoolQ, WinoGrande, COPA, CommonsenseQA and PIQA hide their test labels, so lm-eval scores validation. An index built from test protects the wrong items. These rows index the split in scored_split (shown in bold below).
• Wrong text. Indexing labels or worked solutions, or leaving out the answer choices, changes what counts as a leak. These rows index the text the model is shown.
• Unpinned data. Every row pins a Hub commit, so two indexes built from it agree.
• Short items. An item with fewer than 3 distinct 13-grams can never be flagged by the standard 13-gram rule. short_item_share says how much of each benchmark that is; corpus-assay's exact-match channel (--exact-channel short) covers those items.

Benchmarks

Key	Dataset	Config	Split	Fields	Short items
aime24	HuggingFaceH4/aime_2024		train	problem	0%
aime25	opencompass/AIME2025	data_files	test	question	3%
amc23	math-ai/amc23		test	question	12%
arc	allenai/ai2_arc	ARC-Challenge + ARC-Easy	test	question, choices	49%
arc_challenge	allenai/ai2_arc	ARC-Challenge	test	question, choices	37%
arc_easy	allenai/ai2_arc	ARC-Easy	test	question, choices	55%
boolq	aps/super_glue	boolq	validation	passage, question	1%
commonsenseqa	tau/commonsense_qa		validation	question, choices	95%
copa	aps/super_glue	copa	validation	premise, choice1, choice2	100%
gpqa_diamond	Idavidrein/gpqa (gated)	gpqa_diamond	train	Question, Correct Answer, Incorrect Answer 1, Incorrect Answer 2, Incorrect Answer 3
gsm8k	openai/gsm8k	main	test	question	2%
gsm8k_solutions	openai/gsm8k	main	test	question, answer	0%
hellaswag	Rowan/hellaswag		validation	ctx, endings	0%
humaneval	openai/openai_humaneval		test	prompt	1%
ifeval	google/IFEval		train	prompt	12%
jeebench	daman1209arora/jeebench	data_files	test	question	0%
lambada	EleutherAI/lambada_openai	default	test	text	0%
livecodebench	livecodebench/code_generation_lite	data_files	test	question_content
math500	HuggingFaceH4/MATH-500		test	problem	32%
mbpp	google-research-datasets/mbpp	full	test	text, test_list	4%
mmlu	cais/mmlu	all	test	question, choices	32%
mmlu_pro	TIGER-Lab/MMLU-Pro		test	question, options	8%
openbookqa	allenai/openbookqa	main	test	question_stem, choices	77%
piqa	baber/piqa		validation	goal, sol1, sol2	68%
winogrande	allenai/winogrande	winogrande_xl	validation	sentence, option1, option2	85%

Bold splits are the ones harnesses score because the test labels are hidden. "Short items" is the share of items with fewer than 3 13-grams; blank means not measured.

Use

With corpus-assay, refer to rows by key:

pip install corpus-assay
corpus-assay init-config --benchmarks gsm8k --benchmarks mmlu --out decon.yaml
corpus-assay run --config decon.yaml --input ./shards --out-dir ./scan_out

As data, for your own pipeline:

from datasets import load_dataset

registry = load_dataset("cabbage972/corpus-assay-benchmark-registry", split="train")
row = registry.filter(lambda r: r["key"] == "hellaswag")[0]
print(row["dataset"], row["revision"], row["splits"], row["fields"])

spec_yaml holds each row as a corpus-assay config block.

Columns

Column	Meaning
key	corpus-assay registry key
name	Benchmark name used in reports
dataset, hub_url	Hugging Face dataset id and its page
revision	Pinned Hub commit
config_name / configs	One config, or several indexed as one benchmark
data_files	JSON-encoded data_files override, for datasets whose card lists no splits
splits	Splits indexed
fields	Text fields indexed (each field hashed on its own)
label_field	Label column, used only to check that labels are public; never indexed
scored_split	The split evaluation harnesses score
gated	Needs a Hub login and accepted terms
item_count, short_items, short_item_share	Items, and those too short for the n-gram rule
notes	Caveats for the entry
spec_yaml	The row as a corpus-assay config block
corpus_assay_version	corpus-assay release the rows come from

Links

• Code: https://github.com/mrcabbage972/corpus-assay
• Package: https://pypi.org/project/corpus-assay/
• Registry docs: https://github.com/mrcabbage972/corpus-assay/blob/main/docs/configuration.md#the-built-in-registry
• Software DOI: https://doi.org/10.5281/zenodo.23125328

Citation

@software{may_corpus_assay,
  author    = {May, Victor},
  title     = {corpus-assay: benchmark contamination detection for LLM training corpora},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.23125328},
  url       = {https://github.com/mrcabbage972/corpus-assay}
}

Each benchmark keeps its own license; see its dataset page.