rasyosef/amharic-neural-ir

Official codebase for the ACL 2026 MeLLM Workshop paper "The Multilingual Curse at the Retrieval Layer: Evidence from Amharic" # The Multilingual Curse at the Retrieval Layer: Evidence from Amharic This repository accompanies the ACL 2026 MeLLM Workshop paper **"The Multilingual Curse at the Retrieval Layer: Evidence from Amharic."** It provides notebook and CLI workflows for dense retrieval, late interaction (ColBERT-style), sparse retrieval (SPLADE-style), and cross-encoder reranking in Amharic. **Paper**: The Multilingual Curse at t… **Core artifacts** - **Benchmark**: Amharic Passage Retrieval Dataset V2 with a fixed 90/10 train–test split (68,000 query–passage pairs). - **Model suite**: Amharic-specific checkpoints spanning `dense bi-encoders`, `late-interaction (ColBERT-style)`, `learned sparse retrievers (SPLADE-style)`, and `cross-encoder rerankers`. - **Workflows**: notebook implementations for `preprocessing`, `training`, `evaluation`, `indexing`, `search`, and `RAG` plus CLI/SLURM scripts for selected fine-tuning and evaluation runs. **Hugging Face resources** - **Dataset:** rasyosef/Amharic-Passage-Retrieval-Dataset-V2 - **Monolingual Amharic models:** rasyosef/amharic-neural-ir-models - **Fine-tuned multilingual models:** kiyam/amharic-fine-tuned-multilingual-retrievers **Monolingual Amharic models** - rasyosef/embedding-amharic-base - rasyosef/embedding-amharic-medium - rasyosef/colbert-amharic-base - rasyosef/colbert-amharic-medium - rasyosef/splade-amharic-base - rasyosef/splade-amharic-medium - rasyosef/reranker-amharic-base - rasyosef/reranker-amharic-medium **Amharic-fine-tuned multilingual models** - kiyam/EmbeddingGemma-300M-Amharic — MRR@10: 0.718, NDCG@10: 0.753 - kiyam/Harrier-270M-Amharic — MRR@10: 0.760, NDCG@10: 0.795 **Indexing and Search notebooks** To see the models in action, check out the following `notebooks`. - Amharic Embedding, Reranking & RAG with LlamaIndex > This hands-on guide demonstrates how to use our custom embedding models and cross-encoder rerankers alongside **LlamaIndex** to implement a robust **two-stage retrieval** pipelin …