OribAI — Hausa & Yoruba Language Model

OribAI is an instruction-tuned conversational model fine-tuned for Hausa and Yoruba speakers. It is based on Qwen2.5-14B-Instruct and trained on 27,498 unique Hausa and Yoruba conversational pairs, lexical tasks, and human-annotated instruction data.

Quickstart (Transformers)

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "Alkamal01/oribai-14b-hausa-yoruba-v1"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto")

messages = [
    {"role": "system", "content": "You are OribAI, a helpful Hausa and Yoruba assistant."},
    {"role": "user", "content": "Menene babban birnin Nijeriya?"}
]
input_ids = tokenizer.apply_chat_template(messages, return_tensors="pt").to(model.device)
output = model.generate(input_ids, max_new_tokens=256)
print(tokenizer.decode(output[0], skip_special_tokens=True))

Quickstart (Unsloth)

from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained("Alkamal01/oribai-14b-hausa-yoruba-v1", load_in_4bit=True)
FastLanguageModel.for_inference(model)

Quickstart (llama.cpp / Ollama)

Download the GGUF file from this repo, then:

# llama.cpp
./llama-cli -m oribai-14b-q4_k_m.gguf -p "Menene babban birnin Nijeriya?"

# Ollama (from local GGUF)
ollama create oribai -f Modelfile
ollama run oribai

Example Outputs

Hausa (factual Q&A):

User: Menene babban birnin Nijeriya? OribAI: Abuja

Yoruba (open-ended):

User: Ṣe alaye ìtàn Yoruba fún mi. OribAI: Yorùbá jẹ ènìyàn ọkunrin púpọ̀ tí wọ́n sì ń gbé ní apá ìwọ̀-oòrùn orílẹ̀-èdè Nàìjíríà...

Evaluation

Perplexity measured on 50 samples from CohereForAI/aya_dataset (train split):

Language	Perplexity	Quality Assessment
Yoruba	3.22	Strong — fluent, factually coherent responses
Hausa	62.54	Limited — factual Q&A works, open-ended generation unreliable

⚠️ Note: Hausa generation quality is inconsistent. The model performs better on short factual questions than open-ended or conversational Hausa prompts. Yoruba performance is significantly stronger. Hausa improvement is planned for v2.

Training Data

Dataset	Description
`CohereForAI/aya_dataset`	Hausa + Yoruba multilingual instructions
`CohereForAI/aya_evaluation_suite`	Human-annotated evaluation pairs
`mangaphd/hausa_aug_lex`	14,663 Hausa lexical pairs
`llama-lang-adapt/AfriInstruct-Data`	African instruction data

Training Details

Parameter	Value
Base Model	Qwen2.5-14B-Instruct
Framework	Unsloth + TRL SFTTrainer
LoRA Rank	32
LoRA Alpha	64
Epochs	3
Learning Rate	2e-4
Quantization	4-bit (NF4)

Known Issues

Hausa open-ended generation may hallucinate or go off-topic
Always use a system prompt to establish OribAI identity (see Quickstart)
Short factual answers are more reliable than long-form generation in both languages

Limitations

Model may hallucinate facts, especially for current events
Coverage of Yoruba dialects may be uneven
Not evaluated on formal/legal/medical language use cases
Responses may mix languages occasionally (code-switching)

Citation

@misc{oribai2026,
  author = {Alkamal01},
  title = {OribAI: Hausa and Yoruba Language Model},
  year = {2026},
  publisher = {HuggingFace},
  url = {https://huggingface.co/Alkamal01/oribai-14b-hausa-yoruba-v1}
}