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 https://huggingface.co/meta-llama/Llama-2-7b-hf

2. Download the amharic finetune from huggingface like this:

• git lfs install
• git clone https://huggingface.co/iocuydi/llama-2-amharic-3784m

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