| Language | Country | Speakers |
|---|---|---|
| Hausa | Nigeria/Niger/Chad | 70M+ |
| Yoruba | Nigeria/Benin/Togo | 45M+ |
| Igbo | Nigeria | 30M+ |
| Nigerian Pidgin | West Africa | 75M+ |
| Swahili | Kenya/Tanzania | 200M+ |
| Amharic | Ethiopia | 35M+ |
| Somali | Somalia/Kenya | 20M+ |
| Xhosa | South Africa | 27M+ |
| Zulu | South Africa | 27M+ |
| Shona | Zimbabwe | 15M+ |
| Luganda | Uganda | 8M+ |
| Lingala | DRC/Congo | 70M+ |
| Wolof | Senegal/Gambia | 12M+ |
| Oromo | Ethiopia/Kenya | 40M+ |
| Tigrinya | Eritrea/Ethiopia | 9M+ |
| Rundi | Burundi/DRC | 9M+ |
| Twi | Ghana | 9M+ |
| Fulani | West Africa | 40M+ |
| Malagasy | Madagascar | 25M+ |
| Kinyarwanda | Rwanda | 12M+ |
| Parameter | Value |
|---|---|
| Base Model | Llama 3.2 3B Instruct |
| Training Pairs | 221,697 |
| Training Steps | 1,000 |
| Final Loss | 1.67 |
| Method | QLoRA (r=16) via Unsloth |
| Hardware | NVIDIA Tesla T4 |
| Data Sources | Masakhane, Glot500, CulturaX |
1from transformers import AutoModelForCausalLM, AutoTokenizer
2
3tokenizer = AutoTokenizer.from_pretrained("okaforpascal40/BaobabAI-v0.3")
4model = AutoModelForCausalLM.from_pretrained("okaforpascal40/BaobabAI-v0.3")