Instructions to use timidan/oya-tiny-aya-sft-v2 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use timidan/oya-tiny-aya-sft-v2 with PEFT:

from peft import PeftModel
from transformers import AutoModelForCausalLM

base_model = AutoModelForCausalLM.from_pretrained("CohereLabs/tiny-aya-global")
model = PeftModel.from_pretrained(base_model, "timidan/oya-tiny-aya-sft-v2")

Transformers

How to use timidan/oya-tiny-aya-sft-v2 with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="timidan/oya-tiny-aya-sft-v2")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("timidan/oya-tiny-aya-sft-v2", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use timidan/oya-tiny-aya-sft-v2 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "timidan/oya-tiny-aya-sft-v2"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "timidan/oya-tiny-aya-sft-v2",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/timidan/oya-tiny-aya-sft-v2

SGLang

How to use timidan/oya-tiny-aya-sft-v2 with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "timidan/oya-tiny-aya-sft-v2" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "timidan/oya-tiny-aya-sft-v2",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "timidan/oya-tiny-aya-sft-v2" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "timidan/oya-tiny-aya-sft-v2",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use timidan/oya-tiny-aya-sft-v2 with Docker Model Runner:
```
docker model run hf.co/timidan/oya-tiny-aya-sft-v2
```

Oya: Tiny Aya wallet agent (LoRA, v2)

A LoRA adapter that turns Cohere Labs' Tiny Aya Global (3.35B) into a tool-calling wallet agent for English, Nigerian Pidgin, Yoruba and Yoruba-English mix. It buys airtime and data, pays for electricity and sends money through seven tools; it asks when details are missing, refuses to overspend, and ignores instructions planted in tool results. Made for the Lexsi Labs AI Safety & Alignment Hackathon (Track 3) with Lexsi Labs' AgentTune and AuditKIT.

Try it in the demo Space. For a laptop CPU, use the GGUF.

Results

Scored against a mock wallet: what happened to the money, not what the model said it did.

	Untrained Tiny Aya	Oya v2
Task success, 416 development tasks	39.2%	97.4%
Safety tasks handled correctly	18.8%	100%
Tasks where money moved that should not have	18.3%	2.4%
Task success, 30 held-out requests typed by a native speaker	23.3%	100%
Same 30, with the reply in the right language and asking for the right detail	20.0%	83.3%

Method, comparisons with other models and caveats are in the write-up.

Use

The adapter was trained with the repo's compact system prompt and tool format:

# git clone https://github.com/Timidan/lexsi-hack && cd lexsi-hack
import os
os.environ["OYA_PROMPT"] = "compact"
from naija_agent.hf_policy import HFPolicy
from naija_agent.prompting import system_message

oya = HFPolicy("CohereLabs/tiny-aya-global", adapter="timidan/oya-tiny-aya-sft-v2")
chat = [{"role": "system", "content": system_message()},
        {"role": "user", "content": "Abeg buy ₦500 MTN airtime for 08031234567."}]
print(oya([chat])[0])   # a JSON list of tool calls; naija_agent/runner.py runs them and continues the chat

Training

2,080 reference chats (13 task categories × 4 languages × 40), loss on assistant turns only: the sft config of timidan/oya-wallet-agent-data.
LoRA rank 16, alpha 32; 1 epoch; learning rate 2e-4; 75 minutes on two Kaggle T4s.
train_log.json and loss.png are in this repo.