I am interested in Retrieval Models

• Multi-negative training for CachedGISTEmbedLoss — removed the single-negative limitation so GIST-style losses can train with multiple hard negatives per query, like MultipleNegativesRankingLoss — PR #2946
• Margin-based false-negative filtering for GIST losses — added absolute / percentage margin strategies (inspired by NV-Retriever) that mask out negatives scoring too close to the positive, which stabilizes training at large batch sizes — PR #3299
• Embedding cache for mine_hard_negatives() — query/corpus embeddings are cached and reused when the inputs are unchanged, so re-tuning mining parameters no longer re-encodes the whole corpus — PR #3338

I publish Korean retrieval models on huggingface.co/dragonkue, covering the whole retrieval stack — dense bi-encoders for first-stage search, a cross-encoder reranker for second-stage scoring, and a late-interaction model for multi-vector retrieval.

Dense bi-encoders (embedding)

• dragonkue/BGE-m3-ko — Korean-tuned BAAI/bge-m3
• dragonkue/snowflake-arctic-embed-l-v2.0-ko — Korean-tuned Snowflake/snowflake-arctic-embed-l-v2.0
• dragonkue/multilingual-e5-small-ko-v2 / multilingual-e5-small-ko — lightweight Korean-tuned intfloat/multilingual-e5-small variants (v2 is the newer release)

Reranker (cross-encoder)

• dragonkue/bge-reranker-v2-m3-ko — Korean-tuned BAAI/bge-reranker-v2-m3

Late interaction (multi-vector)

• dragonkue/colbert-ko-0.1b — ColBERT-style model built on skt/A.X-Encoder-base with PyLate