Munyarwanda-AI-v0.2

Munyarwanda-AI v0.2

Munyarwanda-AI v0.2 is a Kinyarwanda-first LLM, planned for QLoRA fine-tuning (r=16, alpha=32) on top of Qwen/Qwen3-0.6B.

Status

Model repo ready for future training — no weights pushed yet. Training will use the notebook Munyarwanda_Train_v02_Final.ipynb (Google Colab T4).

Training Data

Trained on arcange9/Munyarwanda-AI-Dataset (config v0.2) — 5,834 cleaned Kinyarwanda instruction examples (train 5,542 / test 292).

Sources: map-boy/kinyarwanda-dataset-v2, DigitalUmuganda/Monolingual_health_dataset, KoseiUemura/KinNewsClassification.

Mandatory System Prompt

Uri Munyarwanda AI, umufasha w’ubwenge buhangano wibanda ku Kinyarwanda. Igisubizo cyawe kigomba kuba mu Kinyarwanda gisobanutse kandi gisanzwe. Ntukoreshe Icyongereza nk’ururimi rw’igisubizo keretse umukoresha agusabye mu buryo bweruye gusubiza mu Cyongereza cyangwa urundi rurimi.

Quick Start (after training)

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "arcange9/Munyarwanda-AI-v0.2"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto")

prompt = "<|im_start|>system\nUri Munyarwanda AI, umufasha w’ubwenge buhangano wibanda ku Kinyarwanda. Igisubizo cyawe kigomba kuba mu Kinyarwanda gisobanutse kandi gisanzwe. Ntukoreshe Icyongereza nk’ururimi rw’igisubizo keretse umukoresha agusabye mu buryo bweruye gusubiza mu Cyongereza cyangwa urundi rurimi.<|im_end|>\n<|im_start|>user\nSobanura akamaro k'ubuhinzi mu Rwanda<|im_end|>\n<|im_start|>assistant\n"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))