EmbeddingGemma-300M-Amharic

EmbeddingGemma-300M-Amharic

Google EmbeddingGemma-300M fine-tuned on Amharic passage retrieval supervision. This model is presented in the paper:

The Multilingual Curse at the Retrieval Layer: Evidence from Amharic
Yosef Worku Alemneh, Kidist Amde Mekonnen, Maarten de Rijke
The 1st Workshop on Multilinguality in the Era of Large Language Models (MeLLM), ACL 2026

Code: https://github.com/rasyosef/amharic-neural-ir

Results on Amharic Passage Retrieval Dataset V2

Fine-tuning yields a +60.3% relative MRR@10 gain over zero-shot.

Evaluation dataset: rasyosef/Amharic-Passage-Retrieval-Dataset-V2

Usage

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("kiyam/EmbeddingGemma-300M-Amharic")

queries = ["የኢትዮጵያ ዋና ከተማ የትኛው ናት?"]
passages = ["አዲስ አበባ የኢትዮጵያ ዋና ከተማ ናት።"]

query_embeddings = model.encode(queries, normalize_embeddings=True)
passage_embeddings = model.encode(passages, normalize_embeddings=True)

scores = query_embeddings @ passage_embeddings.T

This model uses Matryoshka embeddings — you can truncate to shorter dimensions (e.g. 256) for faster retrieval at a small quality cost:

query_embeddings = model.encode(queries, normalize_embeddings=True)[:, :256]

Training Details

• Base model: google/embeddinggemma-300m (300M parameters)
• Training data: yosefw/amharic-news-retrieval-dataset-v2-with-negatives-V2
• Objective: MultipleNegativesRankingLoss + MatryoshkaLoss (dims: 768, 256)
• Epochs: 6 | Batch size: 32 | Grad. accum.: 4
• Learning rate: 4e-5, cosine schedule | Precision: BF16
• Max sequence length: 512

Citation

@inproceedings{alemneh2026amharicir,
  title     = {The Multilingual Curse at the Retrieval Layer: Evidence from Amharic},
  author    = {Alemneh, Yosef Worku and Mekonnen, Kidist Amde and de Rijke, Maarten},
  booktitle = {Proceedings of the 1st Workshop on Multilinguality in the Era of Large Language Models (MeLLM), ACL 2026},
  year      = {2026},
}