10Academy-FineTuned-LLM/LLM_Finetuning_For_Amharic_Ad_Generation

# LLM Finetuning For Amharic Ad Generation ## About This project aims to fine-tune an llm so that it can understand the Amharic language and create an Advertisement in Amharic given a brand information, product brief. It'll utilize messages exported from 25 publicly available channels to extend the pre-training phase of the model as well as fine-tune the model to generate ads later on. ## Usage At this point in time, you'll need the raw data of the channel messages in a directory named data/raw. Then you can follow the following steps to clean the data and make it appropriate for the model: - `pip install -r requirements.txt` - inside `parse_and_save.ipynb`, run the function `process_raw_data` to get only the necessary data from the raw data which are id, text, date - inside `cleaning.ipynb` run the function `clean_parsed_data` to get the cleaned data which has removed emojis, symbols, newlines, extra spaces To test the inference of the model being used you'll need to follow this steps: 1. Accept Llama2 license on huggingface and download it like this: - git lfs install - git clone huggingface.co 2. Download the amharic finetune from huggingface like this: - git lfs install - git clone iocuydi/llama-2-amharic-378… 3. Clone this github repository 4. Then inside inference/run_inf.py: - change the MAIN_PATH to the path to folder you downloaded from step 1 - change the peft_model to the path you cloned in the step 2 - Go to your llama2 folder(from step 1) and replace the tokenizer related files with the one you find from the 2nd step - set quanitzation=True inside the main function before the load_model function call 5. Finally run the inference/run_inf.py file ## References - Medium Article showcasing how Garri Logistics team fine-tuned Llama to understand Amharic - Key terms and concepts - Fine-tune lama step by step