SlayerLab/slayer-1.07-pretrain

Slayer 1.07 – continued-pretraining data

Tokenized data prepared for continued pretraining of Slayer 1B from the 1.06 checkpoint (step 13848). Public dataset: the exact tokenized files used to train Slayer 1B 1.07: 8 891 005 707 tokens (Polish 56.3 %, English 43.7 %), of which about 7.0 B are new and 1.89 B are a replay of Slayer 1.06 data. The same files, byte for byte, are published in SlayerLab/slayer-1.07-pretrain and Maggio33/slayer-1.07-pretrain. Licenses differ per source; see License before reusing the Polish web part.

Full-size chart

Format

• train.bin in each data directory: little-endian uint16 token ids, tokenizer Slayer v6 (vocab 32 768).
• The exact tokenizer.json is included at the repository root; token ids in every .bin file refer to it. Its SHA256 is 0640d3bd3674d7a6e59c540d945a2ad275a12e1bcaeb6007c37a90751f9815dd. Decode the ids with this tokenizer, not a tokenizer selected by a moving model revision.
• Every document is followed by the end-of-text token 0.
• docs.idx in each data directory: uint64 little-endian, the end offset (in tokens) of each document.

Composition

directory	language	tokens	documents	sources
pl_web/	PL	3 110 430 927	2 445 601	HPLT v3 Polish (files not used by Slayer 1.06)
pl_replay_A/, pl_replay_B/	PL	1 894 295 941	3 409 667	replay subset of the Slayer 1.06 core (982 781 646) and legal (911 514 295) mix
en_mix/	EN	1 650 618 814	1 249 258	FineWeb-Edu, FineMath, open-web-math, DCLM-baseline (subset)
en_fineweb_edu/	EN	2 235 660 025	2 029 345	FineWeb-Edu sample/10BT (4 files)
total		8 891 005 707		9 files, after decontamination and masking

Processing

1. Freshness: English documents whose normalized text is an exact duplicate of a document in the Slayer 1.06 English pool were removed, as were duplicates within the build. Near-duplicates are not detected.
2. Personal data: e-mail addresses, PESEL, NIP, REGON, ID-card numbers and phone numbers are replaced with tags ([EMAIL], [PESEL], [NIP], [REGON], [DOWOD], [TELEFON]), with the same masking function as in Slayer 1.06. In the Polish web part, e-mail addresses had already been replaced with [PII] by the HPLT v3 cleaning; our masking additionally tagged phone, REGON, ID-card, PESEL and NIP numbers in 4 426 documents.
3. Benchmark decontamination: every file was scanned for 13-gram overlap with the test items of the evaluation suite, and documents with any hit were removed. The result was verified document by document, byte for byte, against the input (every kept document is identical, so its scan result is unchanged); files changed after removal by personal-data masking were scanned again with 0 hits. Exact matches of short answers (below 13 words) were counted but not removed. One evaluation set (EWoK) is distributed as a password-protected archive and was not part of the scan. An additional scan covered ARC-Easy English, ARC-Easy-PL EU20 and the English source texts of Belebele (FLORES); it removed another 103 English documents, included in the totals above.
4. Replay: a uniformly random subset (by hash) of documents already seen by Slayer 1.06, cut at document boundaries, minus a list of excluded documents. The Slayer 1.06 core includes the source passage corpora of the PolQA and PoQuAD benchmarks; passages with a 13-gram hit were removed, the rest remain in the replay. Results on these two benchmarks should be read as "source passages in training".

Sources and attribution

• FineWeb-Edu (HuggingFaceFW/fineweb-edu), FineMath (HuggingFaceTB/finemath), open-web-math (open-web-math/open-web-math): ODC-By 1.0. FineWeb-Edu and the datasets derived from Common Crawl are also subject to the Common Crawl Terms of Use. The English subsets come from the public SlayerLab/minimal-en-corpus-5b (revision bb3aec23) and from FineWeb-Edu sample/10BT (revision 87f09149).
• DCLM-baseline (mlfoundations/dclm-baseline-1.0): CC-BY-4.0.
• HPLT v3 Polish: the HPLT compilation is CC0-1.0; the texts are not ours (see License).
• Polish core and legal replay: from the public SlayerLab/slayer-pl-8x3b (revision d50df04a), licenses per source as listed in that dataset card.

License

This dataset is a compilation of sources with different licenses; there is no single blanket license. The license of each part is given in Sources and attribution.

HPLT web part (pl_web/) — legal basis statement: the HPLT project releases its compilation under CC0, but explicitly does not own the underlying texts ("we do not own any of the text"). We therefore do NOT claim CC0 on the web texts themselves and do not grant any license to them. Our publication relies on the text-and-data-mining framework (Directive (EU) 2019/790, Art. 4 — lawful access, opt-out respected at crawl level by HPLT); training use falls under the same TDM basis. Users redistributing or re-publishing the texts are responsible for their own legal review. This is the same basis as for our public SlayerLab/gollem-corpus-16b-pl.