droidnext/gemma_3_270m_kiliki_language

Gemma 3 270M - Kiliki Language Fine-tuned Model

This model is a fine-tuned version of google/gemma-3-270m-it using QLoRA (Quantized Low-Rank Adaptation) for English to Kiliki language translation.

Model Details

• Base Model: google/gemma-3-270m-it
• Fine-tuning Method: QLoRA (4-bit quantization + LoRA adapters)
• Training Dataset: kiliki_dataset_10k.csv (7,528 unique English-Kiliki translation pairs)
• Model Size: Only adapter weights (~few MB) - requires base model for inference

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
from peft import PeftModel
import torch

# Load base model with quantization
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16,
)

model = AutoModelForCausalLM.from_pretrained(
    "google/gemma-3-270m-it",
    quantization_config=bnb_config,
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained("google/gemma-3-270m-it")

# Load QLoRA adapter
model = PeftModel.from_pretrained(model, "droidnext/gemma_3_270m_kiliki_language")

# Generate translation
messages = [{"role": "user", "content": "Hello"}]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=50)
print(tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:]))

Training Details

• Training Dataset: 7,528 English-Kiliki translation pairs
• Training Split: 80% train, 20% test
• Method: QLoRA (4-bit quantization with LoRA rank 64)