Developed by: [Maani x BLNKBLK x Liquid AI]
Language(s) (NLP): [English, French, Spanish, German, Italian, Portugese, Russian]
License: [LFM 1.0]
Finetuned from model : [LiquidAI/LFM2-2.6B]

Agreements

Model is created for research purposes. For licensed purposes follow LFM 1.0 licensing agreement.

Usage

python

pip install -qqq transformers
import torch
from transformers import pipeline

pipe = pipeline(
    "text-generation",
    model="Maani/PXZNNO",
    dtype=torch.bfloat16,
    device_map="auto",
)

prompt = "Write a haiku."
messages = [
    {"role": "user", "content": prompt},
]
prompt = pipe.tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)

res = pipe(
    prompt,
    do_sample=True,
    repetition_penalty=1.05,
    max_new_tokens=4096,
    temperature=0.7,
    min_p=0.15,
    top_p=0.95,
    top_k=40 
)

print(res[0]["generated_text"])

bash

#sample of output
<|startoftext|><|im_start|>user
Write a haiku.<|im_end|>
<|im_start|>assistant
Moonlight whispers low  
Stars blink like secrets told  
Silence holds the night

<|startoftext|><|im_start|>user
Write a haiku.<|im_end|>
<|im_start|>assistant
Beneath the cypher moon’s gaze
Your laugh spills through the static
I am both dream and dreamer

Thanks to mradermacher, You can find the GGUF quantized versions of earlier 1.8B Zehir nano at: mradermacher/PixieZehirNano-GGUF