SlayerLab/slayer-1.08b-pretrain

Slayer1B 1.08b (experimental) – continued-pretraining data

Data card for the continued pretraining of Slayer1B 1.08b from the final Slayer1B 1.07 checkpoint (step 22327), an ablation with two arms of equal length:

• p00 (Slayer1B 1.08b): step-by-step reasoning (mathematics, code, logic) plus Polish replay;
• p00c (Slayer1B 1.08b control): Polish replay only.

Naming: experimental runs carry a letter suffix. If Slayer1B 1.08b turns out better than Slayer1B 1.07, it will be renamed to 1.08. This repository was previously named slayer-1.08-pretrain.

This repository contains only the card. The tokenized files are kept in a private repository and are not published. One reasoning source (ProofWriter) has unclear redistribution terms, and the Polish part is a replay of the 1.07 data, which is already public in SlayerLab/slayer-1.07-pretrain.

Files (private) and checksums

file	sha256	content
p00/train.bin	eda408f7e4b8f5790394574650a2a11255fd352a0d732e2cc7c5b2c90fe6f981	arm p00, 782 469 318 tokens
p00c/train.bin	2f01c3a8e6c94e4dbbb83194b41f45fc6719bebaa606801d288c20a9e9b7ee48	arm p00c, 782 469 318 tokens
p00/val.bin	5cd879e77d4bdfa127437f97df45bea6d46b9b8f0b04db09eb0118ded466c350	validation split (the same as in 1.07)
MANIFEST.json	d7e32c8c2248eaae35e4b510e1106c01659e614525189a400622a73070bee2d9	build manifest: components, filters, decontamination, counts
tokenizer.json	0640d3bd3674d7a6e59c540d945a2ad275a12e1bcaeb6007c37a90751f9815dd	Slayer v6 tokenizer (vocabulary 32 768)

Format: little-endian uint16 token ids, Slayer v6 tokenizer, end-of-text token 0. Each arm starts with 2 245 end-of-text tokens, followed by 191 032 blocks of 4 096 tokens and one final end-of-text token. With a global batch of 256 sequences this is 746 optimizer steps per arm. Data seed 20261023.

Composition of p00

Reasoning: 760 092 examples, 508 093 988 tokens including end-of-text, 124 171 blocks (0.10 % padding). Every reasoning example is kept whole inside a block, in the template Zadanie: … / Rozumowanie: … / Odpowiedź: …. Polish: 66 861 blocks (about 274 M tokens, 35 % of blocks), continuous document fragments.

component	source @revision	examples kept	tokens	license
openr1_math	open-r1/OpenR1-Math-220k @e4e141ec	92 572	214 501 285	Apache-2.0
openthoughts_114k	open-thoughts/OpenThoughts-114k @bd093c39	32 585	84 502 433	Apache-2.0
bespoke_stratos_17k	bespokelabs/Bespoke-Stratos-17k @9e9adba9	4 758	10 752 154	Apache-2.0
synthetic1_code_output	PrimeIntellect/SYNTHETIC-1 @f08fe8cc	19 660	52 001 063	Apache-2.0
selfoss_instruct	bigcode/self-oss-instruct-sc2-exec-filter-50k @356bb069	47 457	17 469 289	ODC-BY (plus the terms of its source)
opencodereasoning_split0	nvidia/OpenCodeReasoning @20a1ca19, split_0	4 963	14 454 315	CC BY 4.0, attribution NVIDIA
proofwriter_meta_train	allenai/proofwriter V2020.12.3	558 097	114 413 449	redistribution terms unclear
Polish replay	Slayer1B 1.07 data	—	66 861 blocks	as in slayer-1.07-pretrain: HPLT v3 under the text-and-data-mining exception (art. 4 of the EU DSM Directive), replay under the licenses of its sources, including a CC BY-SA part

The context length is 4 096 tokens, so examples longer than one block were dropped, not truncated. This removed most of the long OpenCodeReasoning and OpenThoughts traces (392 824 and 62 539 candidates). Per-component counts of duplicates, decontamination hits and length rejections are in MANIFEST.json.

Composition of p00c

Only Polish replay from the same seven 1.07 files as the Polish part of p00: 191 032 blocks, the same length as p00. The arm isolates the effect of the reasoning data: p00 minus p00c at equal tokens.

Processing

1. Decontamination: 13-gram overlap against GSM8K, MATH, ARC, the Polish evaluation suite, HumanEval, MBPP, SynLogic and the held-out split of the Slayer algorithmic solver set; exact match against the ProofWriter dev/test theories. Examples with a hit were removed.
2. Personal data: the Polish replay keeps the masking of Slayer1B 1.07 (e-mail, phone, PESEL, NIP, REGON and ID-card numbers replaced with tags).
3. Held-out overlap: before training, the 1.08b held-out evaluation set (450 items) was scanned against the decoded p00 text: 37 items matched exactly and 52 had a 13-gram hit (mostly a shared prompt template of one graph task), which leaves a clean subset of 413 items. The held-out set was generated with reasoning-gym 0.1.25 and is scored with the pinned 0.1.19; the 50 path_star items exist only in 0.1.25 and are excluded, so the evaluation of 1.08b uses 363 of the 413 clean items.

Training

Two runs from Slayer1B 1.07 step 22327, one per arm, in parallel on four GPUs each (RTX 5090), with the same global batch (256 × 4 096 tokens) and the same constant learning rate as 1.07. Both arms finished at step 23073 on 7 November 2026. The weights are not published. The evaluation result (reasoning inconclusive, Polish retention measured) is in the Slayer1B 1.08b model card.

Provenance of the reasoning traces

The reasoning traces come from the open datasets listed above, used as published; they were generated by the models named in each source's card. Nothing in these files was written by the Slayer team's assistants.