This model was fine-tuned on the Kinyarwanda corpus (kin) of the MasakhaNER dataset. However, we thresholded the number of entity groups per sentence in this dataset to 10 entity groups.

Training procedure

This model was trained on a single NVIDIA P5000 from Paperspace

Hyperparameters

Learning Rate: 5e-5
Batch Size: 32
Maximum Sequence Length: 164
Epochs: 30

Evaluation Data

We evaluated this model on the test split of the Kinyarwandan corpus (kin) present in the MasakhaNER with no thresholding.

Metrics

Precision
Recall
F1-score

Limitations

The size of the pre-trained language model prevents its usage in anything other than research.
Lack of analysis concerning the bias and fairness in these models may make them dangerous if deployed into production system.
The train data is a less populated version of the original dataset in terms of entity groups per sentence. Therefore, this can negatively impact the performance.

Caveats and Recommendations

The topics in the dataset corpus are centered around News. Future training could be done with a more diverse corpus.

Results

Model Name	Precision	Recall	F1-score
roberta-base-kin	76.26	80.58	78.36

Usage

python

from transformers import AutoTokenizer, AutoModelForTokenClassification
from transformers import pipeline

tokenizer = AutoTokenizer.from_pretrained("arnolfokam/roberta-base-kin")
model = AutoModelForTokenClassification.from_pretrained("arnolfokam/roberta-base-kin")

nlp = pipeline("ner", model=model, tokenizer=tokenizer)
example = "Rayon Sports yasinyishije rutahizamu w’Umurundi"

ner_results = nlp(example)
print(ner_results)