SlayerLab/Slayer1B-1.04

Slayer1B 1.04

Decoder-only language model, 1 073 912 851 parameters, Polish-dominant. Slayer1B 1.04 is Slayer1B 1.03 trained further: training resumed from the final checkpoint of 1.03 (step 7150, constant learning rate 6e-4) and continued to step 10010 (10 496 245 760 tokens in total, ≈ 9.8 tokens per parameter).

Like 1.02 and 1.03, 1.04 is a constant-learning-rate state: no learning-rate decay was applied. Pre-training of the Slayer1B line continued from this checkpoint (releases 1.05 at step 12870 and 1.06 at step 13848).

• Previous versions: SlayerLab/Slayer1B-1.03, SlayerLab/Slayer1B-1.02, SlayerLab/Slayer1B-1.01, SlayerLab/Slayer1B-1.00
• Training checkpoints (with optimizer state): SlayerLab/Slayer1B-1.04-checkpoints
• Training curve and training-time evaluation (steps 7160–13848): https://track.fabryka.ai/run/f056bb5c-f1da-56d5-9776-cc33063f10a5

Training

• Same model, tokenizer, corpus order and global batch as 1.03: 256 sequences × 4096 tokens = 1 048 576 tokens per step, constant learning rate 6e-4.
• Steps 7150 → 10010 (2860 steps, 2 998 927 360 tokens) on 4× NVIDIA L40S.
• Data: SlayerLab/slayer-pl-8x3b, continuing the single pass over the corpus started in 1.00.

Training-time evaluation

Measured during training on the released checkpoint (step 10010) with the run's automatic evaluator (one instrument, NVIDIA A100, zero-shot). These are not the lm-evaluation-harness benchmark used in the Results table of 1.00–1.03 and are not directly comparable with it.

metric	1.04 (10010)
BLiMP (EN), accuracy %	74.17
ARC-Easy (EN), accuracy %	50.59
WikiText-2, byte perplexity ↓	2.3078
WikiText-2, token perplexity ↓	31.29
MultiBLiMP (PL), accuracy %	98.87
ARC-Easy (PL, eu20), accuracy %	44.02

The lm-evaluation-harness benchmark (HellaSwag, PIQA, LAMBADA, ARC-Challenge, BoolQ, MMLU) has not been run on 1.04. The benchmark caveat of the 1.03 card (n-gram overlap of the English shared part with BoolQ, MMLU and ARC test items) applies to this release as well: steps up to 7192 (1.00–1.03 and the first 42 steps of the continuation) used the unfiltered data. From step 7193 (pack-04) on, the 5133 documents with n-gram overlap with BoolQ / MMLU / ARC test items are excluded (list sha256 84caed78733c25a7d7bbe5f0c9c1ef1d2bb249793ef559f1b73f4b390995b445).

Files

model.safetensors (weights), config.json, tokenizer.json, load_slayer.py (loader), train_gpt_ref_r23.py and SHA256SUMS. The config.json records the sha256 of the source checkpoint ckpt_10010.pt (04438959d7b317d1a8d972b264e9610a06fd8addde7349c6e9f458ee56dc1e51, also published in the checkpoints repository).

train_gpt_ref_r23.py is the model definition (GPT class) imported by load_slayer.py; trainer_sha256 in config.json refers to this file. The continuation from 1.03 (steps 7150–10010) was trained with train_gpt_ref r24.1 (sha256 881cb055c38b73a24154d96e2359072e9c526f68875c41e0c1b4ed1c4b9eb62c); the architecture is unchanged.