Datasets:

MuseMesh
/

mume-eval-suites

The frozen evaluation sets behind every bits-per-byte number of the Muse Mesh English and Math models (mume-english-125m, mume-math-125m), record for record, including the contamination-clean variants; and training_manifests, the list of source rows each model was trained on (no text), so the training data can be rebuilt from the public sources.

From Muse Mesh (Hugging Face). Part of the Muse Mesh English and Math models (collections). Every training run, including the ones not released, is logged at mume.ai/sansar/runs (moving to mume.ai/lab).

from datasets import load_dataset
gsm = load_dataset("MuseMesh/mume-eval-suites", "gsm8k_test", split="test", revision="v0.1.0")
man = load_dataset("MuseMesh/mume-eval-suites", "training_manifests", split="math_pretrain", revision="v0.1.0")

Evaluation configs

Each config has one split, test, with columns id and text (and url for the OpenWebMath sets). The text is exactly what was scored: models score each set as one stream, records joined by their end-of-document token, and divide the loss by the UTF-8 bytes of text.

config	group	records	MB	what	licence
`fineweb_val`	english	14,983	46.52	FineWeb validation: the first 10,485,760 GPT-2 tokens of llm.c's FineWeb validation shard (14,983 web documents), the set the GPT-2 speedrun scores	ODC-By 1.0 (and the Common Crawl Terms of Use)
`fineweb_val_clean`	english	11,048	28.45	fineweb_val minus the 3,935 documents that share a copied passage with the training data (11,048 documents)	ODC-By 1.0 (and the Common Crawl Terms of Use)
`wikitext103_test`	english	62	1.29	WikiText-103 test, raw, grouped into its 62 articles	CC BY-SA 4.0
`enwik8_test`	english	1	5.00	enwik8 bytes 95,000,000-100,000,000 (the standard test split; Wikipedia XML)	CC BY-SA 3.0 / GFDL (Wikipedia text; see below)
`text8_test`	english	1	5.00	text8 characters 95,000,000-100,000,000 (lowercase letters and spaces)	CC BY-SA 3.0 / GFDL (Wikipedia text; see below)
`owm_val`	math	2,704	20.00	OpenWebMath shard 113 (never trained on), the first 2,704 documents up to 20 MB	ODC-By 1.0 (and the Common Crawl Terms of Use)
`owm_val_clean`	math	1,126	4.78	owm_val minus the 1,578 documents that share a copied passage with the training data (1,126 documents)	ODC-By 1.0 (and the Common Crawl Terms of Use)
`gsm8k_test`	math	1,319	0.65	GSM8K test, 1,319 problems: question, blank line, worked answer (calculator annotations removed)	MIT
`math_test`	math	5,000	3.57	MATH test (Hendrycks), 5,000 problems: problem, blank line, solution	MIT
`math_test_clean`	math	3,754	2.42	math_test minus the 1,246 problems that share a copied passage with the training data (3,754 problems)	MIT

There is no gsm8k_clean: no GSM8K test problem shares a copied passage with the Math training data (0 of 1,319), so gsm8k_test is already clean.

How the clean variants were built

Before training, every evaluation set was checked against the model's whole training slice, GPT-2-paper style (scripts/experts/english/overlap.py): text is lowercased and split into \w+ words; every 8-gram of the evaluation set is hashed; the training slice is streamed once and every hash it contains is marked. A record has a copied passage when 5 or more consecutive 8-grams of it are marked. The _clean config is the set minus every record with a copied passage:

set	8-grams found in training	records with a copied passage	clean config
fineweb_val (against the FineWeb slice)	3.66%	3,935 of 14,983	fineweb_val_clean, 11,048
wikitext103_test	1.67%	18 of 62	none (kept as is)
enwik8_test / text8_test	1.35% / 1.52%	1 of 1 (one 5 MB record)	none
owm_val (against the OpenWebMath slice)	12.31%	1,578 of 2,704	owm_val_clean, 1,126
math_test	4.42%	1,246 of 5,000	math_test_clean, 3,754
gsm8k_test	0.13%	0 of 1,319	not needed

fineweb_val itself stays, as the like-for-like set: llm.c and the GPT-2 speedrun train on FineWeb with the same copies. The full per-record lists are in meta/*_contamination_overlap.json (copied_ids_all).

Sources

fineweb_val: llm.c's FineWeb validation shard (fineweb_val_000000.bin from kjj0/fineweb10B-gpt2, GPT-2 tokens), its first 10,485,760 tokens decoded and split at end-of-text into 14,983 documents. The shard is the head of FineWeb sample-10BT file 000 (llm.c tokenizes in dataset order), which is why the models train on files 001-004 only.
wikitext103_test: Salesforce/wikitext wikitext-103-raw-v1 test, 4,358 raw lines grouped into 62 articles at top-level headings.
enwik8_test, text8_test: the last 5,000,000 bytes / characters of enwik8 / text8 (the standard test split), from mattmahoney.net.
owm_val: open-web-math/open-web-math shard 113 (of 114), documents in file order up to 20,000,000 bytes.
gsm8k_test: openai/gsm8k main test, question + blank line + answer, <<calculator>> annotations removed.
math_test: EleutherAI/hendrycks_math test, seven subjects, problem + blank line + solution.

Revisions of the source repositories are in meta/export_report.json.

training_manifests

One row per source row used. Columns: model_versions, source (Hugging Face dataset), source_revision, source_file (path inside that repository), row (0-based row in that file), doc_id (FineWeb id, OpenWebMath url, empty for GSM8K/MATH), role, parts (records the document became after cutting at 20,000 characters) and val_parts (which of those parts went to the run's 0.5% validation stream instead of training).

split	rows	what
`english_pretrain`	4,180,465	FineWeb documents: mume-english-125m v0.1.0 and v0.2.0; 4,257,824 records after the 20,000-character cut, 21,358 of them in the validation stream
`math_pretrain`	662,940	OpenWebMath pages: mume-math-125m v0.1.0 (and through it v0.1.0-sft); 774,551 records after the 20,000-character cut, 3,769 of them in the validation stream
`math_sft`	14,973	GSM8K and MATH train rows: mume-math-125m v0.1.0-sft; roles: MATH sft_dropped_over_1024_tokens 239, MATH sft_no_boxed_answer 4, MATH sft_train 7,257, GSM8K sft_prompt_shot 3, GSM8K sft_train 7,270, GSM8K sft_val 200

To rebuild a pretraining slice: read each source_file at source_revision, take the listed rows in order, cut documents longer than 20,000 characters at the last space before the limit and at least half way in (else the last newline there, else at 20,000; the next part starts after stripping leading whitespace), drop the val_parts, tokenize, and put the end-of-document token after every part. Training read random 1,024-token windows from that stream, so order does not matter. The manifests were checked against our slice files (same ids or word counts, in order) and against the preparation records (record and validation counts).

Licence

Each config keeps the licence of its source; this repository adds no restriction. The manifests and the metadata files are released under ODC-By 1.0 (they index FineWeb, OpenWebMath, GSM8K and MATH).

licence	configs
ODC-By 1.0, plus the Common Crawl Terms of Use	fineweb_val, fineweb_val_clean, owm_val, owm_val_clean, training_manifests
CC BY-SA 4.0	wikitext103_test
CC BY-SA 3.0 / GFDL	enwik8_test, text8_test
MIT	gsm8k_test, math_test, math_test_clean

enwik8 and text8 are distributed by Matt Mahoney without a licence statement of their own. Their text is the English Wikipedia dump of 2006-03-03, which was published under the GFDL and has been available under CC BY-SA 3.0 since Wikipedia's 2009 relicensing; we redistribute the 5 MB test slices on those terms, with attribution to Wikipedia's contributors. The FineWeb validation shard was re-hosted as GPT-2 tokens in kjj0/fineweb10B-gpt2 (MIT); the text is FineWeb's (ODC-By 1.0).

Attribution

data	Hugging Face	reference	licence
FineWeb	HuggingFaceFW/fineweb	Penedo et al. 2024, The FineWeb Datasets, arXiv:2406.17557	ODC-By 1.0; use is also subject to the Common Crawl Terms of Use
OpenWebMath	open-web-math/open-web-math	Paster et al. 2023, OpenWebMath, arXiv:2310.06786	ODC-By 1.0; use is also subject to the Common Crawl Terms of Use
GSM8K	openai/gsm8k	Cobbe et al. 2021, Training Verifiers to Solve Math Word Problems, arXiv:2110.14168	MIT
MATH	EleutherAI/hendrycks_math	Hendrycks et al. 2021, Measuring Mathematical Problem Solving With the MATH Dataset, arXiv:2103.03874	MIT
WikiText-103	Salesforce/wikitext	Merity et al. 2016, Pointer Sentinel Mixture Models, arXiv:1609.07843	CC BY-SA 4.0 (per its dataset card)
enwik8 / text8	n/a	Matt Mahoney, mattmahoney.net/dc/textdata.html; the first 10^8 bytes of the English Wikipedia dump of 2006-03-03	no licence stated by the distributor; the text is English Wikipedia (GFDL at the time, CC BY-SA 3.0 since the 2009 relicensing)