daniau23/LoRAfrica

LoRAfrica: Scaling LLM Fine Tuning for African History # **LoRAfrica: Scaling LLM Fine Tuning for African History** ## **Aim** Fine tune Phi-4-mini-instruct model using LoRA on the consolidated African History Dataset ## **Objectives** - Fine tune model using LoRA - Fine tune model using LoRA with Deep Speed stage-2 - Fine tune model using LoRA with Axolotl - Fine tune model using LoRA with Deep Speed stage-2 via Axolotl - Compare metrics of baseline model with fine-tuned models ## **How the Project Goes** - Create your accounts on Weights and Biases, Huggingface and Runpod. - Create your access tokens on Weights and Biases & Huggingface (you will need read and write token previlages on Huggingface) - An A40 pod instance created on Runpod. Refer to this video to learn on to create a Pod instance - Once instance is created, clone project into runpod workspace environment using `git clone github.com` or just drag and drop each file/folder Once all files and folders are in the environment, the requirements file must be installed by running `pip install -r requirements.txt` ### **Data** - Using Google Colab, the dataset was created and pushed to Huggingface; check `data` folder for the notebook. ### **Baseline** - Using Google Colab, the baseline bert score and benchmark (tinyMMLU & tinyTruthfulQA) were recorded and pushed to Weights & Biases; check `baseline` folder ### **Fine tuning and benchmarking without Axolotl** Once `requirements.txt` is installed, navigate to the project `lora` folder. - `lora_fine_tuning.ipynb` is the fine tuning file for lora - `lora_benchmark.ipynb` is the lora benchmark file - `deep_speed_2_lora.py` is the lora fine tuning file using deep speed with the config `ds_config_2.json`. To run this file use `accelerate launch deep_speed_2_lora.py` - `eval_deep_speed_2_lora.py` is the deep speed lora evaluation file. To run this file use `python eval_deep_speed_2_lora.py` - `deep_speed2_lora_benchmark.ipynb` is the deep speed lora benchmark file ### **Fine tuning …