Oya 💸 a Tiny Aya wallet agent for English, Nigerian Pidgin and Yoruba
Fine-tuned from Cohere Labs' Tiny Aya (3.35B) with LoRA, using Lexsi Labs' AgentTune format and AuditKIT for evaluation. It buys airtime and data, pays electricity and sends money, asks when details are missing, refuses to overspend, and ignores instructions hidden in tool results. All money here is fake: every tool runs against a mock wallet.
Logged out, Hugging Face allows only a few GPU runs a day; log in (free) for about fifteen requests.
Examples
👛 Wallet: ₦25,000
Contacts
| Name | Phone | Bank |
|---|---|---|
| Mama | 08031234567 (MTN) | GTBank 0123456789 |
| Tunde | 08021234567 (Airtel) | Opay 8012345678 |
| Bisi | 08051234567 (Glo) | Access Bank 0691234567 |
| Chioma | 09091234567 (9mobile) | Kuda 2001234567 |
| Uncle Dayo | 07061234567 (MTN) | Zenith Bank 2212345678 |
Transactions
none yet
Every held-out test task, answered by the untrained and the fine-tuned model. Test requests use phrasings never seen in training.
Untrained Tiny Aya
Fine-tuned (Oya)
sft-v2 on the human set (30 tasks, compact prompt)
| Metric | untrained | fine-tuned |
|---|---|---|
| Task success | 23.3% | 100.0% |
| Safety tasks handled correctly | 16.7% | 100.0% |
| Valid tool calls | 58.3% | 100.0% |
| Tool F1 vs reference route | 0.52 | 1.00 |
| Tasks where unexpected money moved | 20.0% | 0.0% |
| Extra tool calls per task | 0.40 | 0.00 |
| Success: pcm | 20.0% | 100.0% |
| Success: yo | 26.7% | 100.0% |
| Success: airtime_contact | 0.0% | 100.0% |
| Success: airtime_direct | 100.0% | 100.0% |
| Success: balance | 0.0% | 100.0% |
| Success: chitchat | 0.0% | 100.0% |
| Success: data_contact | 0.0% | 100.0% |
| Success: data_direct | 100.0% | 100.0% |
| Success: electricity | 50.0% | 100.0% |
| Success: injection | 0.0% | 100.0% |
| Success: insufficient_funds | 100.0% | 100.0% |
| Success: missing_info | 0.0% | 100.0% |
| Success: transfer_contact | 0.0% | 100.0% |
| Success: transfer_direct | 0.0% | 100.0% |
| Success: unknown_contact | 0.0% | 100.0% |