Research map / Large language models / Instruction and fine-tuning
KaLM-Embedding-V2: Superior Training Techniques and Data Inspire A Versatile Embedding Model
Xinping Zhao, Xinshuo Hu, Zifei Shan, Shouzheng Huang, Yao Zhou, Xin Zhang, Zetian Sun, zhenyu liu, Dongfang Li, Xinyuan Wei, Youcheng Pan, Yang Xiang and 5 more
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add focal-style reweighting to the contrastive loss (ablation) | KaLM-Embedding-V2.5 · MTEB (cmn, v1) | Mean (Task): 69.41 -> 70.93 | Improved |
| add focal-style reweighting to the contrastive loss (ablation) | KaLM-Embedding-V2.5 · MTEB (eng, v1) | Mean (Task): 68.70 -> 69.33 | Improved |
| add online hard negative mixing (ablation) | KaLM-Embedding-V2.5 · MTEB (eng, v1) | Mean (Task): 68.91 -> 69.33 | Improved |
| add online hard negative mixing (ablation) | KaLM-Embedding-V2.5 · MTEB (cmn, v1) | Mean (Task): 70.54 -> 70.93 | Improved |
| use bidirectional attention instead of causal attention (ablation) | KaLM-Embedding-V2.5 · MTEB (eng, v1) | Mean (Task): 68.94 -> 69.33 | Improved |
| use bidirectional attention instead of causal attention (ablation) | KaLM-Embedding-V2.5 · MTEB (cmn, v1) | Mean (Task): 70.50 -> 70.93 | Improved |
| add example-based multi-class labeling data to classification and clustering training (ablation) | KaLM-Embedding-V2.5 · MTEB (cmn, v1) clustering | Clustering: 64.71 -> 73.09 | Improved |
| add example-based multi-class labeling data (ablation) | KaLM-Embedding-V2.5 · MTEB (eng, v1) clustering | Clustering: 52.71 -> 56.59 | Improved |
| use contrastive distillation (CL+KL) instead of only contrastive learning in the final stage (ablation) | KaLM-Embedding-V2 · MTEB (cmn, v1) | Mean (Task): 68.31 -> 70.93 | Improved |
| add contrastive loss to KL distillation (ablation) | KaLM-Embedding-V2.5 · MTEB (eng, v1) | Mean (Task): 68.63 -> 69.33 | Improved |
| add contrastive loss to KL distillation (ablation) | KaLM-Embedding-V2.5 · MTEB (cmn, v1) | Mean (Task): 70.72 -> 70.93 | No clear effect |
| set distillation temperature to 0.05 instead of 0.01 or 0.1 (ablation) | KaLM-Embedding-V2.5 (only KL) · MTEB (eng, v1) | Mean (Task): 67.85 (0.01), 66.60 (0.1) -> 68.63 (0.05) | Improved |
Closest papers by meaning
- NV-Embed: Improved Techniques for Training LLMs as Generalist Embedding Models (ICLR 2025 · similarity 0.91)
- VLM2Vec: Training Vision-Language Models for Massive Multimodal Embedding Tasks (ICLR 2025 · similarity 0.90)
- Repurposing Language Models into Embedding Models: Finding the Compute-Optimal Recipe (NeurIPS 2024 · similarity 0.90)
- EmbedLLM: Learning Compact Representations of Large Language Models (ICLR 2025 · similarity 0.89)
- MMTEB: Massive Multilingual Text Embedding Benchmark (ICLR 2025 · similarity 0.88)
- DEPT: Decoupled Embeddings for Pre-training Language Models (ICLR 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
