kinyarwanda-lora-afriqueqwen4b

Kinyarwanda LoRA Adapter

A LoRA adapter fine-tuned for Kinyarwanda instruction-following, trained on top of the frozen base model McGill-NLP/AfriqueQwen-4B.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model_name = "McGill-NLP/AfriqueQwen-4B"
adapter_repo = "Leonard250/kinyarwanda-lora-afriqueqwen4b"

tokenizer = AutoTokenizer.from_pretrained(base_model_name)
base_model = AutoModelForCausalLM.from_pretrained(base_model_name)
model = PeftModel.from_pretrained(base_model, adapter_repo)

prompt = "Amabwiriza: Sobanura icyo Irembo ari cyo.\nIgisubizo:"
inputs = tokenizer(prompt, return_tensors="pt")
output = model.generate(**inputs, max_new_tokens=100)
print(tokenizer.decode(output[0], skip_special_tokens=True))

Design notes

• Base model weights are untouched - this adapter only adds trained LoRA matrices (~0.8% of total params).
• Intended for adapter-swapping: attach only when Kinyarwanda is detected; run the base model with no adapter for other languages.
• Trained on a cleaned/filtered subset of ChrisToukmaji/kinyarwanda_instruction_tuning.