Hugging Face's logo

KaLM-Embedding
/
KaLM-Reranker-V1-Large-R2

Text Ranking
sentence-transformers
Safetensors
Transformers
multilingual
t5gemma2
text2text-generation
reranker
encoder-decoder
FBNL
matryoshka
retrieval
RAG
Model card Files
xet
 Community
Yuki131 commited on
Commit
a862dd8
·
verified ·
1 Parent(s): 592cb47

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +11 -0
README.md CHANGED
@@ -59,6 +59,17 @@ library_name: sentence-transformers
59
  | Multi-domain ranking | Strong | **Improved on BEIR for all three sizes** |
60
  | Multilingual ranking | Limited | **Substantially improved across MIRACL's 18 languages** |
61
 
 
 
 
 
 
 
 
 
 
 
 
62
  ### KaLM-Reranker-V1-R2 vs. the original KaLM-Reranker-V1
63
  All results below use the same default compression ratio, `r=4`. KaLM-Reranker-V1 results are taken from the previous model cards; KaLM-Reranker-V1-R2 results are from the updated paper.
64
 
 
59
  | Multi-domain ranking | Strong | **Improved on BEIR for all three sizes** |
60
  | Multilingual ranking | Limited | **Substantially improved across MIRACL's 18 languages** |
61
 
62
+ ### R2 Training Checkpoints (Stages 1–3)
63
+
64
+ We release the checkpoints from the three-stage training pipeline described in the third version of our paper. Stage 1 uses supervised fine-tuning; Stage 2 produces two checkpoints through soft-label distillation; and Stage 3 combines them through model soup to produce the final R2 models.
65
+
66
+ | Stage | Checkpoint | Nano | Small | Large |
67
+ | :--- | :--- | :--- | :--- | :--- |
68
+ | Stage 1 | Supervised fine-tuning | [KaLM-Reranker-V1-Nano-R2-Stage1](https://huggingface.co/Yuki131/KaLM-Reranker-V1-Nano-R2-Stage1) | [KaLM-Reranker-V1-Small-R2-Stage1](https://huggingface.co/Yuki131/KaLM-Reranker-V1-Small-R2-Stage1) | [KaLM-Reranker-V1-Large-R2-Stage1](https://huggingface.co/Yuki131/KaLM-Reranker-V1-Large-R2-Stage1) |
69
+ | Stage 2 | Distillation (`r64-a32`) | [KaLM-Reranker-V1-Nano-R2-Stage2-r64-a32](https://huggingface.co/Yuki131/KaLM-Reranker-V1-Nano-R2-Stage2-r64-a32) | [KaLM-Reranker-V1-Small-R2-Stage2-r64-a32](https://huggingface.co/Yuki131/KaLM-Reranker-V1-Small-R2-Stage2-r64-a32) | [KaLM-Reranker-V1-Large-R2-Stage2-r64-a32](https://huggingface.co/Yuki131/KaLM-Reranker-V1-Large-R2-Stage2-r64-a32) |
70
+ | Stage 2 | Distillation (`r96-a48`) | [KaLM-Reranker-V1-Nano-R2-Stage2-r96-a48](https://huggingface.co/Yuki131/KaLM-Reranker-V1-Nano-R2-Stage2-r96-a48) | [KaLM-Reranker-V1-Small-R2-Stage2-r96-a48](https://huggingface.co/Yuki131/KaLM-Reranker-V1-Small-R2-Stage2-r96-a48) | [KaLM-Reranker-V1-Large-R2-Stage2-r96-a48](https://huggingface.co/Yuki131/KaLM-Reranker-V1-Large-R2-Stage2-r96-a48) |
71
+ | Stage 3 | Final R2 model | [KaLM-Reranker-V1-Nano-R2](https://huggingface.co/KaLM-Embedding/KaLM-Reranker-V1-Nano-R2) | [KaLM-Reranker-V1-Small-R2](https://huggingface.co/KaLM-Embedding/KaLM-Reranker-V1-Small-R2) | [KaLM-Reranker-V1-Large-R2](https://huggingface.co/KaLM-Embedding/KaLM-Reranker-V1-Large-R2) |
72
+
73
  ### KaLM-Reranker-V1-R2 vs. the original KaLM-Reranker-V1
74
  All results below use the same default compression ratio, `r=4`. KaLM-Reranker-V1 results are taken from the previous model cards; KaLM-Reranker-V1-R2 results are from the updated paper.
75