praise-jaravani/African-Language-FineTuning

# CSC5035Z Assignment 2 — Fine-Tuning Language Models on African Language NLP Tasks **Course:** CSC5035Z Natural Language Processing, UCT 2026 **Student number:** JRVPRA001 **Language:** Yoruba (`yor`) **Extension:** B — Vocabulary Adaptation --- ## Project Description This project fine-tunes `jhu-clsp/mmBERT-small` (a compact multilingual BERT encoder) on two Yoruba NLP tasks from the AfroBench benchmark: 1. **MasakhaNews** — news topic classification (9 categories, metric: macro-F1) 2. **MasakhaNER 2.0** — named entity recognition (PER, ORG, LOC, DATE, metric: seqeval span-F1) **Extension B (Vocabulary Adaptation):** The mmBERT-small tokeniser is extended with ~1,000 Yoruba-specific BPE tokens, each initialised as the mean of its constituent subword embeddings from the original model. Both tasks are then re-fine-tuned on this extended model to measure the impact on tokeniser fertility and downstream performance. --- ## Quickest Path — Inspect Pre-computed Results (No Training Required) All result files are included in the submission. To view tables and figures without running any training: **Option A — Jupyter (local):** ```bash # 1. Create and activate a virtual environment python -m venv venv # Windows: venv\Scripts\activate # macOS / Linux: source venv/bin/activate # 2. Install dependencies pip install -r requirements.txt pip install jupyter # 3. Launch Jupyter and open the results notebook jupyter notebook notebooks/results_analysis.ipynb ``` Then run all cells (Kernel → Restart & Run All). **Option B — VS Code:** Open the project folder in VS Code, select the Python interpreter from your virtual environment, then open `notebooks/results_analysis.ipynb` and click **Run All**. **Option C — Google Colab:** Upload `notebooks/results_analysis.ipynb` to Colab and upload the `results/` folder to your Drive at `MyDrive/csc5035z-a2/results/`. Run all cells. --- ## Full Pipeline — Reproduce Training from Scratch Requires a GPU (tested on NVIDIA T4). …