Large Language Model for Amharic

Using text corpora from Amharic wikipedia and Ethiopian News Headlines, we build AMBERT a language embedding bootstrapped off BERT and Google Translate. See the emedding for details.

The resulting embedding can be used via the AmBert class to build various NLP applications.

am2en is a basic example of a model using this embedding for an amharic to english translation application.

Setup

BERT_BASE environment variable, which is used in bert_embedding, should point to the BERT installation directory.

AM_LLM environment variable, which is used in ambert_embedding, should point to the am-llm repository's directory i.e. this directory.

Add $AM_LLM to your PYTHONPATH environment variable.

E.g.

export BERT_BASE=~/bert_base
export AM_LLM=~/src/am-llm
export PYTHONPATH=$AM_LLM:$PYTHONPATH