NIRVLab/splade_morpheme

SPLADE + Morpheme Tokenizer — EViRAL

Cross-lingual retrieval model:

Ede query → Vietnamese passage retrieval

Architecture

• Backbone: 6-layer Transformer
• Retrieval head: SPLADE sparse expansion
• Tokenizer: Morpheme tokenizer v4

Files

File	Description
mlm.pt	MLM pre-trained encoder
align.pt	Cross-lingual aligned encoder
finetune.pt	Fine-tuned retrieval encoder

Notes

This model replaces vanilla dense retrieval with SPLADE sparse lexical expansion while keeping the same Transformer backbone.