zeeshatron-cpu/Somali-Classifier-

A Llama based binary name classifier for accurate classification of the Somali race # Somali Name Classifier A binary classifier that distinguishes Somali names from non-Somali names, trained using LLaMA 3.1 8B Instruct with LoRA fine-tuning. ## Repository Structure ``` ├── README.md ├── train_llama.py # LLaMA 3.1 training script (Google Colab) ├── dupefinder.py # Remove duplicate names from dataset └── synthmaker.py # Generate synthetic name pairs for augmentation ``` --- ## Scripts ### `train_llama.py` — Model Training Trains a LLaMA 3.1 8B Instruct binary classifier with LoRA. Designed to run on **Google Colab** with GPU. **Prerequisites:** - Google Colab with a GPU runtime - HuggingFace account with access granted to `meta-llama/Llama-3.1-8B-Instruct` **Setup:** 1. Upload `train_dataset.csv` and `eval_dataset.csv` to Google Drive at `/MyDrive/Datasets/` 2. Open `train_llama.py` in Colab and run all cells **Dataset format:** ```csv Full Name,label AHMED MOHAMED,1 JOHN SMITH,0 FATIMA HUSSEIN,1 EMILY JOHNSON,0 ``` - `label = 1` → Somali name - `label = 0` → Non-Somali name **Output:** Model and tokenizer saved to `/content/drive/MyDrive/Llama3_Binary_LoRA` --- ### `dupefinder.py` — Duplicate Cleaner Scans `somali_classifier_full_dataset.csv` for duplicate names, reports a breakdown by label, and writes two output files. ```bash python dupefinder.py ``` **Output files:** - `duplicates_found.csv` — all duplicate rows - `cleaned_dataset.csv` — deduplicated dataset ready for training --- ### `synthmaker.py` — Synthetic Name Generator Generates synthetic name combinations from `gooddata.csv` by randomly mixing first and last name parts within each label group (Somali and non-Somali separately), then removes any names that already exist in the original dataset. ```bash python synthmaker.py ``` **Input:** `gooddata.csv` with `Full Name` and `label` columns **Output:** `synthetic_dataset.csv` --- ## Training Configuration | Parameter | Value | |-----------|-------| | Base model | `meta-llama/Llama-3.1-8B-Instruct` …