Kasawa — SmolLM2-135M Fine-tuned on Twi

kasawa is a compact Twi language model capable of generating coherent twi text as casual language model. It should be noted that this model serves as a base model and it has not undergone intruction tuning.

Built on SmolLM2-135M and trained on the Pristine Twi Dataset — ~999k rows of clean, naturally sounding Twi text spanning four styles: narrative, dialogue, monologue, and storytelling, grounded in real Ghanaian news topics and named entities.

Try it the yourself in the demo.

Intended Use

This model is released for research and non-profit use only. The primary goal is to lower the barrier for experimentation in the Twi/Akan NLP space — particularly as a base for instruction-tuned models targeting more advanced tasks such as summarization, question answering, and dialogue in Twi.

It can also be very useful as a backbone for LLM-based text-to-speech (TTS) and automatic speech recognition (ASR) systems, as well as machine translation (MT) models involving Twi.

Quick Start

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline

model = AutoModelForCausalLM.from_pretrained(
    "ghananlpcommunity/kasawa",
    torch_dtype=torch.bfloat16,
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("ghananlpcommunity/kasawa")
gen = pipeline("text-generation", model=model, tokenizer=tokenizer)

prompt = "Ɔpɛnimaa bosome, 2025,"
out = gen(prompt, max_new_tokens=80, do_sample=True, temperature=0.8, top_p=0.9)
print(out[0]["generated_text"])

Training Details

Base model	SmolLM2-135M
Dataset	Pristine Twi (~999k rows, ~250M tokens)
Epochs	~2
Hardware	1× 80GB GPU

Citation & Community

Created by Mich-Seth Owusu for the Ghana NLP Community.
If you build on this model, please credit the original work and consider sharing your results back with the community.

License: Research and non-profit use only.