ML / NLP Engineer (10 Academy capstone)
Fine-tuning Llama 2 for Amharic Text Generation
Fine-tuned Llama 2 to enable quality embeddings and text generation in Amharic, then used it inside a RAG-based ad-copy builder for the Ethiopian market.
Project gallery
Use the side arrows to browse; click an image for full size.
Hugging Face-hosted model + retrieval pipeline for Amharic ad generation.
Fine-tuning Llama 2 for Amharic Text Generation
A Hugging Face-hosted Llama 2 fine-tune for Amharic, plus a RAG pipeline that uses it to generate ad copy for the Ethiopian market, bringing a low-resource language into a usable generative workflow.
The problem
Most open LLMs handle Amharic poorly: tokenization is inefficient and generation quality is low, so off-the-shelf models can't produce usable Amharic embeddings or copy.
What I built
- Fine-tuned Llama 2 on Amharic data to improve embedding quality and text generation for the language.
- Published the model on Hugging Face so it's reusable, not trapped in a notebook.
- A RAG-based ad-copy builder that retrieves relevant context and generates on-brand Amharic ad copy.
How I measured quality
I compared embedding and generation quality before and after fine-tuning on Amharic samples; the Medium write-up details the data prep, training setup, and the quality gains that made the downstream RAG viable.
Stack
Llama 2, Hugging Face Transformers, PyTorch, LangChain, RAG.