Logo Lanfrica

rasyosef/text-embedding-models-training

Domain:

natural language processing

Record type:

modelsoftware
Creator:
ras
Host:
Notebooks to train and evaluate Amharic Text Embedding Models based on BERT and RoBERTa for Passage Retrieval # Training Amharic Text Embedding Models for Passage Retrieval This repo contains code for training Amharic Text Embedding models based on 3 Amharic Encoder Base models - RoBERTa Base Amharic - RoBERTa Medium Amharic - BERT Medium Amharic We also Evaluate the highest ranking multilingual embedding models form the Massive Text Embedding Benchmark (MTEB) Leaderboard using the same test set as our embedding models. # Results Our largest embedding model, RoBERTa-Base-Amharic-Embed beats all of the multilingual embedding models on the MRR@10, NDCG@10 and Recall metrics while having 1/5th of their param count. |Model | Params | MRR@10 | NDCG@10 | Recall@10 | Recall@50 | Recall@100 | |------|--------|--------|---------|-----------|-----------|------------| |gte-modernbert-base | 149M | 0.019 | 0.022 | 0.030 | 0.054 | 0.065 | |gte-multilingual-base | 305M | 0.649 | 0.684 | 0.794 | 0.876 | 0.904 | |multilingual-e5-large-instruct | 560M | 0.713 | 0.747 | 0.853 | 0.924 | 0.946 | |snowflake-arctic-embed-l-v2.0 | 568M | 0.719 | 0.755 | 0.868 | 0.941 | 0.957 | |BERT-Medium-Amharic-Embed | 40M | 0.657 | 0.696 | 0.817 | 0.916 | 0.945 | |RoBERTa-Medium-Amharic-Embed | 42M | 0.707 | 0.744 | 0.861 | 0.941 | 0.963 | |RoBERTa-Base-Amharic-Embed | 110M | **0.755** |**0.790** | **0.897** | **0.957** | **0.971** | # Code The training and eval code can be found in the `notebooks` folder. # Embedding Models You can download and use our Amharic Text Embedding models from HuggingFace and they are fully compatable with the Sentence Transformers Library - RoBERTa-Base-Amharic-Embed: rasyosef/roberta-amharic-te… - RoBERTa-Medium-Amharic-Embed: rasyosef/roberta-amharic-te… - BERT-Medium-Amharic-Embed: rasyosef/bert-amharic-text-…