MicroLlama-v2

<p align="center"> <img src="banner.jpg" alt="MicroLlama Banner" width="100%"> </p>

MicroLlama-v2 (45M Parameters)

The second generation of the MicroLlama series. This is an ultra-compact, 45-million parameter language model trained from scratch and fine-tuned using Supervised Fine-Tuning (SFT) on the clean Alpaca dataset.

Model Features

• Size: ~45M parameters — runs lightning-fast even on a potato.
• Vocabulary: Custom LLaMA-based tokenizer.
• Grammar Over Mind: Thanks to SFT, the model has pristine English grammar and strict formatting, though its world-knowledge is hilariously limited by its scale.
• Lore: Has a strange obsession with cats, parks, and existential AI humor.

Example Dialogue

User: Tell me a joke.
Assistant: One example of an AI system could be that use AI to learn more about the potential consequences or disadvantages of humanity and what is expected? It can also cause data in both languages.

User: Write a sentence about a dog.
Assistant: A cake has an easy way to stay in the top of the mountain, but there are plenty of food that can have many benefits and resources available on people's favorite?

How To Use

from transformers import AutoModelForCausalLM, AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("ViorikaAI-org/MicroLlama-v2")
model = AutoModelForCausalLM.from_pretrained("ViorikaAI-org/MicroLlama-v2")

prompt = "user: How are you?\nassistant: <s> "
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=50)
print(tokenizer.decode(outputs[0]))