OribAI: A Specialized Hausa & Yoruba Conversational Model
OribAI is an instruction-tuned conversational model developed by Alkamal01, built upon the Qwen2.5-14B-Instruct architecture. With 7.6 billion parameters and a 32768-token context length, it is uniquely fine-tuned for Hausa and Yoruba speakers.
Key Capabilities & Features
- Bilingual Focus: Optimized for conversational interactions in Hausa and Yoruba.
- Instruction-Tuned: Trained on 27,498 unique conversational pairs, lexical tasks, and human-annotated instruction data.
- Strong Yoruba Performance: Achieves a perplexity of 3.22 on Yoruba samples from
CohereForAI/aya_dataset, delivering fluent and factually coherent responses. - Hausa Factual Q&A: While open-ended Hausa generation can be inconsistent (perplexity 62.54), it performs reliably for short, factual questions.
- Base Model: Leverages the robust capabilities of Qwen2.5-14B-Instruct.
When to Use This Model
- Conversational AI in Yoruba: Ideal for chatbots and virtual assistants interacting in Yoruba.
- Factual Information Retrieval: Suitable for answering direct questions in both Hausa and Yoruba.
- Language Adaptation: Developers seeking a model specifically trained on African instruction data for these languages.
Limitations to Consider
- Hausa open-ended generation may exhibit hallucinations or go off-topic.
- Responses might occasionally mix languages (code-switching).
- Not evaluated for formal, legal, or medical use cases, and may hallucinate facts, especially for current events.