MarathiXRetrieve : Cross-Lingual Information Retrieval (CLIR) Adaptation for Marathi-English RAG Systems

Abstract This technical report presents MarathiXRetrieve, an empirical study on mitigating the cross-lingual performance penalty in dense retrieval systems for low-resource languages. We develop and evaluate a complete end-to-end pipeline that synthetically generates aligned evaluation datasets and applies contrastive fine-tuning to the BAAI/bge-m3 multilingual embedding model, targeting Marathi-to-English document retrieval scenarios. Our core hypothesis — that targeted contrastive fine-tuning on synthetically aligned, article-aware language pairs can structurally reorganize the joint vector space to mitigate cross-lingual penalties without catastrophic forgetting — is validated through comprehensive experimentation. The fine-tuned model achieves a +2.84% absolute improvement in cross-lingual Top-1 retrieval accuracy and a +3.91% improvement in NDCG@10. Crucially, the performance gap between English and Marathi Top-1 retrieval is compressed from 7.93% to 4.17%, effectively reducing the cross-lingual penalty by 47.4%, while monolingual English performance degrades by only -0.92%, confirming preservation of baseline knowledge. The report covers the full system architecture, including (1) synthetic Q&A generation using Qwen2.5-7B-Instruct-AWQ on vLLM, (2) neural translation via sarvam-translate, (3) vector indexing with ChromaDB and BGE-M3, (4) contrastive optimization using MultipleNegativesRankingLoss, and (5) evaluation via pytrec_eval with a novel 3-tier graded relevance framework. The entire pipeline was executed on a single NVIDIA RTX 3090 (24GB VRAM) at an infrastructure cost of approximately $0.60 USD. All model weights and synthetic datasets are openly available on the Hugging Face Hub. Keywords: Cross-lingual information retrieval, low-resource NLP, Marathi language, dense passage retrieval, contrastive fine-tuning, retrieval-augmented generation (RAG), BGE-M3, sentence transformers Related Identifiers Source Code Repository: github.com Fine-Tuned Model: huggingface.co License This technical report and the associated codebase are released under the MIT License.