Datasets:
chunk stringlengths 3 3 | slot float64 0 16 ⌀ | step_from float64 0 80k ⌀ | step_to float64 5k 85k ⌀ | status stringclasses 3
values | positions_capacity float64 5.35B 5.35B ⌀ | updates_done float64 0 5k ⌀ | positions_trained float64 0 5.24B ⌀ | pl_percent float64 20 30 ⌀ | positions_by_source stringlengths 45 98 | chunk_sha256 stringlengths 64 64 ⌀ |
|---|---|---|---|---|---|---|---|---|---|---|
c01 | 0 | 0 | 5,000 | trained | 5,347,737,900 | 5,000 | 5,242,880,000 | 20 | fineweb-edu=4278190159; gollem-web=1069547741 | 3341080241aeabd1e03c77530a6d735cb9501ecd788fba60ea065a944cd09ee1 |
c02 | 1 | 5,000 | 10,000 | trained | 5,347,740,964 | 5,000 | 5,242,880,000 | 20 | fineweb-edu=4278192922; gollem-web=1069548042 | 0b064039733dc7d61684a8def5859fc166f7d34331ee2659a97cbad96950f23b |
c03 | 2 | 10,000 | 15,000 | trained | 5,347,740,627 | 5,000 | 5,242,880,000 | 20 | fineweb-edu=4278191040; finepdfs-edu-pol=1069549587 | cf477bd496fcb430d7e5e9980edfab3f43899b11bb246fcf6258f04d91003821 |
c04 | 3 | 15,000 | 20,000 | trained | 5,347,852,265 | 5,000 | 5,242,880,000 | 20 | fineweb-edu=4278192783; finepdfs-edu-pol=1069659482 | f2eaf21d7482a9646c32d580244ef581af5e080c86bd646e096a3a8cf4d444f7 |
c05 | 4 | 20,000 | 25,000 | trained | 5,347,778,178 | 5,000 | 5,242,880,000 | 20 | fineweb-edu=4278190586; finepdfs-edu-pol=1069587592 | 6e457d8a084f7eef24cd95be52456f920a81376283c39b6e3a55aa63a1f9b9e6 |
c06 | 5 | 25,000 | 30,000 | trained | 5,347,899,736 | 5,000 | 5,242,880,000 | 20 | fineweb-edu=4278191437; finepdfs-edu-pol=107114594; gollem-web=962593705 | 9d5df696f146dd0546858d635c4ed2518dc8d5ebc002d3fc80b9f7ef849dbd9e |
c07 | 6 | 30,000 | 35,000 | trained | 5,347,742,870 | 5,000 | 5,242,880,000 | 20 | fineweb-edu=4278192273; gollem-web=510761546; fineweb-2-pol=558789051 | 7fc4c309dbdbf181b32e1fc40f9c933a57d0ce3e25f948d8f518ff2b835d6a0b |
c08 | 7 | 35,000 | 40,000 | trained | 5,347,738,610 | 5,000 | 5,242,880,000 | 20 | fineweb-edu=4278190760; fineweb-2-pol=1069547850 | b221fdee490addcfb6dc85e26b4a693957bc4bfc5156c4381987fce556c600a1 |
c09 | 8 | 40,000 | 45,000 | trained | 5,347,739,414 | 5,000 | 5,242,880,000 | 20 | fineweb-edu=4278191718; fineweb-2-pol=1069547696 | 78ceb2368476c485d2ac65c1a246f83ce116e8a33a45666096e531ca11d239e0 |
c10 | 9 | 45,000 | 50,000 | trained | 5,347,738,989 | 5,000 | 5,242,880,000 | 20 | fineweb-edu=4278191330; fineweb-2-pol=1069547659 | d71a86a610cea64d8cf681134835b5cdeec530f9c2fa45a58e4e0dfac161168c |
c11 | 10 | 50,000 | 55,000 | trained | 5,347,741,526 | 5,000 | 5,242,880,000 | 20 | fineweb-edu=4278190947; fineweb-2-pol=1069550579 | 71fc63acb7b518e5fbb491019713b90b008ac391ecf1ea8c691f2e9ee5f5a000 |
c12 | 11 | 55,000 | 60,000 | trained | 5,347,740,406 | 5,000 | 5,242,880,000 | 20 | fineweb-edu=4278192742; fineweb-2-pol=1069547664 | cccb709a3e0b49e089fdd98839c52cafdc11318cfb9ca6cbed355aa152301cab |
c13 | 12 | 60,000 | 65,000 | trained | 5,347,739,613 | 5,000 | 5,242,880,000 | 20 | fineweb-edu=4278191088; fineweb-2-pol=1069548525 | 7d15fc51d0d2657bba9149331b87b733038e610c85766a1135d4a756fa5edf95 |
c14 | 13 | 65,000 | 70,000 | trained | 5,347,740,692 | 5,000 | 5,242,880,000 | 20 | fineweb-edu=4278192539; fineweb-2-pol=1069548153 | b55c0f5136db4b31f37a2bcceb50852c27be9e70492fb4b552d8cd595a2c6a87 |
c15 | 14 | 70,000 | 75,000 | queued | 5,347,741,480 | 0 | 0 | 20 | fineweb-edu=4278193706; fineweb-2-pol=1069547774 | d697afb1a6d2966bf711bc9f0513206600a6d7072e9811c477a1f913f8f72967 |
c16 | 15 | 75,000 | 80,000 | queued | 5,347,740,239 | 0 | 0 | 30 | fineweb-edu=3743418406; fineweb-2-pol=1604321833 | e798de58c0cec3250370f49fb8a231ecf5b94a9a094e5541b36d2ce76a94c662 |
c17 | 16 | 80,000 | 85,000 | queued | 5,347,744,730 | 0 | 0 | 30 | fineweb-edu=3743416494; fineweb2-hq-pol=561513500; finepdfs-pol=240654060; fineweb-2-pol=802160676 | 091d558556f55cfa3911ba758e61ecf1bde2e9b8687e5d4ea26e4c0c31d6f7fc |
c18 | null | null | null | planned | null | null | null | null | not built; drawn from the unassigned pool in components.csv | null |
c19 | null | null | null | planned | null | null | null | null | not built; drawn from the unassigned pool in components.csv | null |
c20 | null | null | null | planned | null | null | null | null | not built; drawn from the unassigned pool in components.csv | null |
Slayer1B Dataset
Arkadiusz Słota, Kacper Wikieł, Anna Sieczka, Piotr Styła, Aleksander Ogrodzki, Paweł Puzio, Marcin Kuśmierski / Fabryka AI (SlayerLab)
A single, versioned inventory of the data used to train the Slayer1B model line (1.00 to 1.09), with the training state of every part: what the model has already been trained on, what is built and waiting in the queue, and what is planned or proposed.
This first release is a manifest, not a text dump. It contains no document text and no token files. Every upstream file behind the 1.09 chunks is identified by repository, full revision, file path and SHA-256, so those inputs can be downloaded and checked from their original publishers; earlier model versions are identified by dataset and revision only. The exact processed subset of each chunk cannot yet be rebuilt from this manifest alone (see How the 1.09 chunks were built). Text of our own components will be added source by source, only after a privacy and licence review of the exact file being published (see Text release plan).
State at this release
State frozen at update 70,000 of Slayer1B 1.09. The full training checkpoint at that update has SHA-256 4efefde3dd55bb368cae821dcaab0dca47e93bb5aaa66f2529df570cdcdb70f2 (training event log; checkpoint written 21 November 2026, 12:35 UTC). All statuses and trained counts below refer to that checkpoint, not to the live run.
Model weights at update 70,000 are published in SlayerLab/Slayer1B-1.09-checkpoints as wagi_70000.pt (SHA-256 9ed73494fc139422f5314f72b4f970242454a7ab9d8caceee6b52fe5bdfc40c0). That file holds the weights only; it is not the full training checkpoint named above.
| Token positions | Note | |
|---|---|---|
| Inherited from Slayer1B 1.00–1.08 (lower bound) | 18,863,257,682 | disjoint by source and file; see lineage |
| 1.09 chunks c01–c17, built | 90,911,898,239 | input capacity of the queue, not consumed exposure |
| 1.09 trained so far | 73,400,320,000 | 70,000 updates × 1,048,576 positions |
| Lineage total at this checkpoint | 92,148,950,554 | 18,863,257,682 + updates × 1,048,576 − 114,627,128 |
The −114,627,128 term is a measured correction for duplicates found in chunk c01. Lineage totals are a lower bound that is disjoint by source and file; they are not a count of globally content-unique tokens. Recomputation after a preemption repeats updates already counted once; repeated work adds no positions here.
A token position is one token of the Slayer v6 tokenizer (vocabulary 32,768, SHA-256 0640d3bd3674d7a6e59c540d945a2ad275a12e1bcaeb6007c37a90751f9815dd), including end-of-document tokens.
Status vocabulary
| Status | Meaning |
|---|---|
trained |
All 5,000 updates of the chunk's slot have run. Within a slot, training windows of 4,096 positions are drawn without replacement, and 5,000 updates × 256 windows cover about 98% of a chunk's positions once. This describes the slot only; it says nothing about unique content, documents, or repeated computation after a preemption. |
trained to step K |
The slot is in progress; updates up to K have run. |
queued |
Built, byte-checked and in the training queue; not yet reached. The run may stop before a queued chunk is used. |
planned |
Upstream files are pinned; the chunk is not built yet. |
proposed |
Submitted in a pull request; not yet reviewed or built. |
1.09 chunks
Each chunk is the input for one 5,000-update slot. Source allocation is 80% English / 20% Polish for c01–c15 and 70% / 30% from c16 (updates 75,001 onward). The PL share column is the share of Polish-source positions in each chunk's built input capacity; it is not the language share of everything the model has seen (for the earlier versions see lineage). In chunks, positions_capacity is the size of the built chunk and positions_trained is updates_done × 1,048,576 at the frozen checkpoint; the two are different measures. Status lives only in chunks; every row of components belongs to one chunk and shares its status. Rows of components with chunk unassigned are pinned upstream files reserved for the planned chunks c18–c20, not yet allocated to a particular chunk.
| Chunk | Updates | Status | Capacity (positions) | Trained (positions) | PL share (1.09 input capacity) |
|---|---|---|---|---|---|
| c01 | 1–5,000 | trained |
5,347,737,900 | 5,242,880,000 | 20.00% |
| c02 | 5,001–10,000 | trained |
5,347,740,964 | 5,242,880,000 | 20.00% |
| c03 | 10,001–15,000 | trained |
5,347,740,627 | 5,242,880,000 | 20.00% |
| c04 | 15,001–20,000 | trained |
5,347,852,265 | 5,242,880,000 | 20.00% |
| c05 | 20,001–25,000 | trained |
5,347,778,178 | 5,242,880,000 | 20.00% |
| c06 | 25,001–30,000 | trained |
5,347,899,736 | 5,242,880,000 | 20.00% |
| c07 | 30,001–35,000 | trained |
5,347,742,870 | 5,242,880,000 | 20.00% |
| c08 | 35,001–40,000 | trained |
5,347,738,610 | 5,242,880,000 | 20.00% |
| c09 | 40,001–45,000 | trained |
5,347,739,414 | 5,242,880,000 | 20.00% |
| c10 | 45,001–50,000 | trained |
5,347,738,989 | 5,242,880,000 | 20.00% |
| c11 | 50,001–55,000 | trained |
5,347,741,526 | 5,242,880,000 | 20.00% |
| c12 | 55,001–60,000 | trained |
5,347,740,406 | 5,242,880,000 | 20.00% |
| c13 | 60,001–65,000 | trained |
5,347,739,613 | 5,242,880,000 | 20.00% |
| c14 | 65,001–70,000 | trained |
5,347,740,692 | 5,242,880,000 | 20.00% |
| c15 | 70,001–75,000 | queued |
5,347,741,480 | 0 | 20.00% |
| c16 | 75,001–80,000 | queued |
5,347,740,239 | 0 | 30.00% |
| c17 | 80,001–85,000 | queued |
5,347,744,730 | 0 | 30.00% |
| c18 | not scheduled | planned |
— | — | — |
| c19 | not scheduled | planned |
— | — | — |
| c20 | not scheduled | planned |
— | — | — |
Prepared replacement for c17. A rebuilt variant of c17 (c17g3; tokenized train.bin SHA-256 adcf3ca50f8ed681f289e1a47b763b087f767596b3293338c1ec3b85be34c49d) was built and byte-checked on 21 November 2026. At this release it is outside the training rotation: the model has not been trained on it, the active training queue still lists the original c17, and chunks and components describe that original c17. If the replacement enters the run, a later release will record it.
Sources
| Source | Language | Upstream (revision) | Licence | 1.09 chunks | 1.09 positions |
|---|---|---|---|---|---|
| fineweb-edu | en | HuggingFaceFW/fineweb-edu (87f09149) |
odc-by | c01–c17 (17) | 71,659,710,930 |
| fineweb-2-pol | pl | HuggingFaceFW/fineweb-2 (af9c1333) |
odc-by | c07–c17 (11) | 11,521,657,460 |
| fineweb2-hq-pol | pl | epfml/FineWeb2-HQ (c0c06e94) |
odc-by | c17 | 561,513,500 |
| finepdfs-edu-pol | pl | HuggingFaceFW/finepdfs-edu (9cfabe21) |
odc-by | c03–c06 (4) | 3,315,911,255 |
| finepdfs-pol | pl | HuggingFaceFW/finepdfs (220bac3a) |
odc-by | c17 | 240,654,060 |
| gollem-web | pl | SlayerLab/gollem-corpus-16b-pl (17f9d37a) |
other | c01–c07 (4) | 3,612,451,034 |
This table covers the whole built 1.09 queue c01–c17. At update 70,000 the model had been trained on c01–c14 only: fineweb2-hq-pol and finepdfs-pol occur only in c17 and had not yet been seen, and part of the fineweb-edu and fineweb-2-pol positions listed here lie in the queued chunks c15–c17.
Licence values are the license field of each upstream dataset card at the pinned revision. ODC-By requires attribution to the upstream dataset; the underlying web pages and PDFs remain subject to Common Crawl terms of use and to the rights of their authors. GoLLeM web is Polish web text from HPLT v3; for its licence basis see the GoLLeM card (License section). A licence for the model does not grant rights to redistribute training texts.
Earlier model versions (lineage)
| Model versions | Data | New positions (lower bound) | PL share | PL basis (denominator, positions) | Text release |
|---|---|---|---|---|---|
| 1.00–1.06 | Slayer PL 8×3B: eight Polish training packs with a shared core, a legal layer, a disjoint HPLT v3 web slice per pack and an English top-up; its card counts about 24 billion cl100k tokens, and its built training input is about 14.5 billion positions with the Slayer v6 tokenizer | 14,520,680,448 | 81.80% | built train-input capacity of the 8 retained v6 manifests, validation removed (14,524,133,431) | pending privacy review |
| 1.07 | Slayer 1.07 pretraining mix: continued pretraining from 1.06 with new Polish web text (HPLT v3), English text (FineWeb-Edu, FineMath, open-web-math, a DCLM subset) and a replay of the 1.00–1.06 core and legal data | 3,110,230,927 | 56.29% | reconstructed whole-phase sampled target positions, including replay (8,890,875,904) | pending privacy review |
| 1.08 | PL Logika 2B, English Technical 1B, replay of the 1.07 mix | 1,232,346,307 | 73.85% | reconstructed whole-phase sampled target positions, including replay (1,792,016,384) | linked upstream; replay part pending privacy review |
1.00 started from random weights; every later version continues the previous one. The 1.08 figure counts only positions that the reproduced 1.08 sampler drew from data new in 1.08. Earlier versions are listed by dataset and revision; file-level pins are not given for them. Slayer PL 8×3B (1.00–1.06, replayed in 1.07) includes a Polish web layer from HPLT v3; its card lists that layer as CC0-1.0, which covers the HPLT packaging, and, as with the 1.09 web sources, the underlying web pages remain subject to the rights of their authors. Filtering later data does not change already published checkpoints; this card does not measure memorisation or extraction.
PL share of earlier versions. Language is assigned per source component, not detected document by document. Each row has its own denominator, given in the table, and the denominators differ from the lower-bound counts in the third column: for 1.00–1.06 the share is measured on the built training input of the eight retained manifests (14,524,133,431 positions), not on the 14,520,680,448 positions sampled; for 1.07 and 1.08 it covers every sampled target position of the whole phase, replay included, not only the positions new in that version. For Slayer PL 8×3B itself, its data contract (dataset.json) gives 89.66% Polish for the physical pool and its card gives 90% as the sampling plan; both are counted with the cl100k tokenizer and are not converted here. The overall Polish share across the whole lineage is not measured in this release: the shares above use different denominators and must not be combined or weighted by the lower-bound counts.
How the 1.09 chunks were built
For each upstream file: tokenization with the v6 tokenizer; whole-document removal when a personal-data detector finds an e-mail address, a phone number, a Polish identifier (PESEL, NIP, REGON, identity-card number), a bank-account or payment-card number, or a US social-security number; exact deduplication within the source, across chunks and against the data of Slayer1B 1.00–1.08; benchmark decontamination (13-gram and 8-gram template scans against an extended reference set of evaluation benchmarks, with positive controls, from c03 onward, where for the FinePDFs-edu component of c03 this is not yet confirmed; c01 and part of c02 were scanned only against an earlier reference list that did not contain these sets); and a held-out sample of documents (configured at 1%) kept outside training. In components, input_positions counts the tokenized upstream file before these steps and positions_in_chunk counts what was selected into the chunk after them; document-level filter accounting is not part of this release.
Dropping whole documents with a detected identifier lowers privacy risk; it does not prove that no personal data remains. Removal and selection lists are not part of this release, so the exact document subset of a chunk cannot yet be rebuilt from this repository alone; the upstream files and the per-chunk SHA-256 of the tokenized result are given so that a rebuild can be checked.
Text release plan
- Now: this manifest. Public upstream sources are referenced, not copied.
- Next, source by source: text of our own components, each after an independent privacy check on the exact published file (the published bytes must be the checked bytes) and a licence review. The table above changes from "pending" to a link when a component is released.
- Sources whose terms do not allow redistribution stay as references only.
Contributing
New data enters as proposed through a pull request to manifest/proposed.csv; see CONTRIBUTING.md. Proposals are reviewed for licence, personal data and overlap with existing data before they can be built into a chunk. A source accepted after review is added to this manifest as a further chunk, so the dataset can keep growing; a private source is still listed, marked private, and no text or files from it are published.
Citation and attribution
If you use this manifest, cite the Slayer1B model card and attribute each upstream dataset you download according to its own card.
- Downloads last month
- 84