sayikhushhal5-sketch/low-resource-hausa-sentiment

# Low-Resource Hausa Sentiment Analysis with Transfer Learning End-to-end deep learning project for three-class sentiment classification (negative / neutral / positive) on Hausa text. Built on `xlm-roberta-base` fine-tuned on the AfriSenti Hausa benchmark and evaluated out-of-domain on NollySenti Hausa movie reviews. --- ## Live links | Resource | URL | |---|---| | Live Gradio demo (Hugging Face Space) | huggingface.co | | Fine-tuned model weights (Hugging Face Hub) | sayikhushhal/xlm-r-hausa-se… | | Colab notebook (training + evaluation + deployment) | colab.research.google.com | The Hugging Face Space serves the live demo with five pre-loaded NollySenti example chips — anyone can try the model in the browser in under 10 seconds without cloning or installing anything. --- ## Repository layout ``` . ├── README.md — this file ├── notebook/ │ └── SayiKhushhalGadde_A00074661_MS4.ipynb — full training, evaluation, deployment pipeline ├── gradio-app/ │ ├── app.py — Gradio interface (loads model from HF Hub) │ ├── requirements.txt — runtime deps for the Space │ └── README.md — Space metadata and one-line summary ├── report/ │ └── SayiKhushhalGadde_A00074661_MS4.pdf — IEEE-format final report └── requirements.txt — deps for re-running the notebook ``` --- ## Reproducing the results — Colab path (recommended) The notebook was developed and tested on Google Colab with an NVIDIA T4 GPU. The full top-to-bottom run takes approximately **25 minutes** and trains 9 model variants plus the deployment pipeline. 1. Open the Colab notebook URL above (or upload `notebook/SayiKhushhalGadde_A00074661_MS4.ipynb` to your own Colab). 2. **Runtime → Change runt …