Hailay/FT_EXLMR

---

1. Model Description

Hailay/FT_EXLMR is a fine-tuned version of the EXLMR model, designed specifically for sentiment analysis and text classification tasks in low-resource African languages such as Tigrinya, Amharic, and Oromo. This model leverages the architecture of EXLMR but has been further fine-tuned to improve its performance on multilingual tasks, especially for languages not widely represented in existing NLP models. The model was trained using the AfriSent-Semeval-2023 dataset, a benchmark dataset for African languages, which is publicly available on GitHub:AfriSent-Semeval-2023 GitHub Repository

2.Intended Use

This model is ideal for: Researchers and developers who are working on multilingual sentiment analysis in African languages. Applications that require text classification in low-resource languages. It is designed specifically for tasks such as: Sentiment analysis Text classification

Note: Without further fine-tuning, the model is unsuitable for tasks like machine translation or named entity recognition.

3.Training Data

The Hailay/FT_EXLMR model was trained using the dataset from the SemEval 2023 Shared Task 12: Sentiment Analysis in African Languages (AfriSenti-SemEval). This dataset comprises sentiment-labeled text from 14 African languages:

1. Algerian Arabic (arq) - Algeria
2. Amharic (ama) - Ethiopia
3. Hausa (hau) - Nigeria
4. Igbo (ibo) - Nigeria
5. Kinyarwanda (kin) - Rwanda
6. Moroccan Arabic/Darija (ary) - Morocco
7. Mozambique Portuguese (pt-MZ) - Mozambique
8. Nigerian Pidgin (pcm) - Nigeria
9. Oromo (orm) - Ethiopia
10. Swahili (swa) - Kenya/Tanzania
11. Tigrinya (tir) - Ethiopia
12. Twi (twi) - Ghana
13. Xithonga (tso) - Mozambique
14. Yoruba (yor) - Nigeria

The dataset covers diverse data for training multilingual models like Hailay/FT_EXLMR We access the dataset from AfriSent-Semeval-2023 GitHub Repository. The Hailay/FT_EXLMR model was trained using the following configuration: Epochs: 3 Learning Rate: 1e-5 Optimizer: AdamW Batch Size: 16

4. Evaluation

The model was evaluated using accuracy and loss as the primary metrics. The results are as follows:

Accuracy: Achieved strong performance on Tigrinya, Amharic, Afar, and Oromo text classification and sentiment analysis tasks.

Loss: Loss values showed steady convergence during the 3 epochs of training, reflecting a well-calibrated model. The evaluation was carried out on the test set provided in the AfriSent-Semeval-2023 GitHub Repository dataset.