oolel-lit-gemma

Oolel-lit-gemma

Oolel-lit-gemma is a fine-tuned version of Gemma-3-270m-it for the Wolof language. It is part of our Oolel family of compact, on-device Wolof language models developed.

The model was trained using supervised fine-tuning (SFT) on synthetic data distilled from our larger Oolel-7B models via Oolel-translator.

Usage

Quick start with pipeline

from transformers import pipeline

generator = pipeline(
    "text-generation",
    model="soynade-research/oolel-lit-gemma",
    device="cuda",
)

messages = [{"role": "user", "content": "Translate to Wolof: The president is 45 years old."}]

output = generator(messages, max_new_tokens=256, return_full_text=False)
print(output["generated_text"])

With AutoModel for more control

from transformers import AutoTokenizer, Gemma3ForCausalLM
import torch

model_id = "soynade-research/oolel-lit-gemma"


model = Gemma3ForCausalLM.from_pretrained(
    model_id
).eval()

tokenizer = AutoTokenizer.from_pretrained(model_id)

messages = [
    [
        {
            "role": "system",
            "content": [{"type": "text", "text": "You're a Wolof AI assistant. Please always provide detailed and useful answers to the user queries."},]
        },
        {
            "role": "user",
            "content": [{"type": "text", "text": "Translate to Wolof: The president is 45 years old."},]
        },
    ],
]

inputs = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    tokenize=True,
    return_dict=True,
    return_tensors="pt",
).to(model.device).to(torch.bfloat16)


with torch.inference_mode():
    outputs = model.generate(**inputs, max_new_tokens=256,         
                             do_sample=True,
                             temperature=0.7,
                             top_p=0.9,)

outputs = tokenizer.batch_decode(outputs)

Training

The training code and configuration are available at soynade-research/oolel-trainer.

Limitations

• Primarily optimized for Wolof; performance on other languages may vary
• As a 270M parameter model, it may struggle with complex tasks
• Outputs should be verified by a native Wolof speaker for critical applications