Gemmaiku-3-270m-it is an experimental fine-tuned version of Google's Gemma-3-270m-it, specialized in generating traditional 5-7-5 syllable haikus on any topic.

It was trained locally using LoRA on Apple Silicon using the Apple MLX framework.

[!NOTE] Syllable Strictness: The model is trained to target a strict 5-7-5 structure. However, due to its compact size (270M parameters), it may occasionally deviate slightly on complex or out-of-distribution prompts. For best results, keep temperatures low (0.1 - 0.3) and use the correct chat template.

Usage

python

from mlx_lm import load, generate
from mlx_lm.sample_utils import make_sampler

# Load the model
model, tokenizer = load("vi-c0de/gemmaiku-3-270m-it-experimental")

# Keep temperature low for strict syllable adherence
sampler = make_sampler(temp=0.3)

# Format using the tokenizer chat template (ensures correct bos token)
messages = [{"role": "user", "content": "I want to become a hardware engineer"}]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)

# Generate response
response = generate(model, tokenizer, prompt=prompt, sampler=sampler)
print(response.split("<end_of_turn>")[0].strip())