KiloNovaSynth-12B

Where matter itself is squeezed to the edge of meaning, two dying stars spiral inward.
The silence breaks with a final collision, forging new worlds in the ashes.
Light blooms briefly, soft but searing, as gravity sings in frequencies never heard.
This is not fire, but memory collapsed into mass.

🔧 Recommended Sampling Settings:

Temperature: 0.75 to 1.25
Min P: 0.035
Context Length: Stable at 12k tokens, with possible support for extended contexts

💬 Prompt Format

Supports ChatML style messages. Example:

<|im_start|>user
Your question here.
<|im_end|>
<|im_start|>assistant

KiloNovaSynth-12B is a merge of the following models using LazyMergekit:

🧩 Configuration

merge_method: dare_ties
base_model:
 model: DreadPoor/Irix-12B-Model_Stock
 name: Irix
models:
 - model: yamatazen/LorablatedStock-12B
 name: Lorablated
 parameters:
 weight: 0.4
 density: 0.7
 - model: yamatazen/EtherealAurora-12B-v2
 name: Aurora
 parameters:
 weight: 0.3
 density: 0.6
parameters:
 normalize: true
 int8_mask: true
 anneal_factor: 0.2
dtype: bfloat16

layer_parameters:
 - filter: "attn"
 sources:
 - model: Irix
 weight: 0.8
 - model: Aurora
 weight: 0.15
 - model: Lorablated
 weight: 0.05
 
 - filter: "mlp"
 sources:
 - model: Lorablated
 weight: 0.6
 - model: Aurora
 weight: 0.3
 - model: Irix
 weight: 0.1

💻 Usage

!pip install -qU transformers accelerate

from transformers import AutoTokenizer
import transformers
import torch

model = "Marcjoni/KiloNovaSynth-12B"
messages = [{"role": "user", "content": "What is a large language model?"}]

tokenizer = AutoTokenizer.from_pretrained(model)
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
pipeline = transformers.pipeline(
 "text-generation",
 model=model,
 torch_dtype=torch.float16,
 device_map="auto",
)

outputs = pipeline(prompt, max_new_tokens=256, do_sample=True, temperature=1, top_k=0, top_p=1)
print(outputs[0]["generated_text"])