
Organization	NULLXES
Contact	ceo@nullxes.com
Architecture	`MurzikForCausalLM` (custom, not a fork)
Total params	~13B
Precision	bf16
HF repo	`MagistrTheOne/murzik-15b-init` (this page)

Current checkpoint


Stage	Pre-training (first run)
Steps	1500
Data	Wikipedia (en/ru/de/es/fr/zh/uk) + Murzik identity corpus
Seq length	2048
Tokens seen	~49M (~0.37 epoch)
Chat / instructions	Not yet — SFT is the next stage

Weights in this repo are updated in place (random init → PT → later SFT).
The repo name stays murzik-15b-init; only the README and files change per stage.

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "MagistrTheOne/murzik-15b-init"

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    trust_remote_code=True,
    torch_dtype="auto",
    device_map="auto",
)

Chat template (after SFT)

Template name in LlamaFactory: murzik

<|murzik|><|system|>
{system}<|end|>
<|user|>
{user}<|end|>
<|assistant|>
{assistant}<|end|>

Roadmap

Stage	Status
Random init	done
Pre-training	done (first run, 1500 steps)
SFT (identity + Aya)	next
MoE 32B	separate line

License

Proprietary — NULLXES. Weights are published for research and integration testing.
Commercial use requires written permission: ceo@nullxes.com.

Citation

bibtex

@misc{murzik15b_init2026,
  title        = {NULLXES MURZIK-15B},
  author       = {NULLXES},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/MagistrTheOne/murzik-15b-init}},
  contact      = {ceo@nullxes.com}
}