What is Kestrel?

Kestrel is a lightweight, local language model for Mauritian Creole created as the capstone of my Master's Postgraduate Degree in Computer Science at University College London (UCL). Built on the IBM Granite 4.1 family of open-weight base models, Kestrel aims to democratise artificial intelligence for lower-resource languages - those that have much smaller publicly-available corpora of data than major world languages such as English and French.

The project investigates the fine-tuning of Granite for French- and English-Mauritian Creole translation, using BLEU and chrF++ scores to provide a quantitative analysis of results.

Developed by Felix Everett for MSc Computer Science

Supervised by Prof. Dean Mohamedally

UCL Computer Science

Results (chrF++ scores)

Results were measured with BLEU and chrF++, with the latter detailed below, as research (Mathur et al., 2020) has suggested that it is less prone to skew at the lower scores typical in low-resource model fine-tuning. Note that these are machine-translation evaluation metrics, which, while quantitative, do not provide conclusive insight into the translation quality of these models. Human analysis is recommended to further this research. A full list of recommended steps to further this investigation can be found in section 5.4 of the accompanying paper for these models.

English model

Model EN-KM KM-EN Average
3b-base 12.62 20.52 16.89
3b-kestrel-en 35.97 42.73 39.51
8b-base 13.97 27.86 21.10
8b-kestrel-en 37.30 41.95 39.74

French model

Model FR-KM KM-FR Average
3b-base 17.00 18.78 17.94
3b-kestrel-fr 34.67 34.24 34.46
8b-base 16.91 28.30 23.04
8b-kestrel-fr 36.28 36.99 36.70
Downloads last month
55
GGUF
Model size
3B params
Architecture
granite
Hardware compatibility
Log In to add your hardware

4-bit

16-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for felixreverett/kestrel

Quantized
(66)
this model

Dataset used to train felixreverett/kestrel