51 lines
1.4 KiB
Markdown
51 lines
1.4 KiB
Markdown
---
|
|
language:
|
|
- ko
|
|
license: apache-2.0
|
|
tags:
|
|
- sentence-transformers
|
|
- sentence-similarity
|
|
- transformers
|
|
---
|
|
|
|
## PwC-Embedding-expr
|
|
|
|
We trained the **PwC-Embedding-expr** model on top of the [multilingual-e5-large-instruct](https://huggingface.co/intfloat/multilingual-e5-large-instruct) embedding model.
|
|
To enhance performance in Korean, we applied our curated augmentation to STS datasets and fine-tuned the E5 model using a carefully balanced ratio across datasets.
|
|
|
|
> ⚠️ This is an experimental model and is under continuous development.
|
|
|
|
### To-do
|
|
- [x] MTEB Leaderboard
|
|
- [ ] Technical Report
|
|
|
|
|
|
## MTEB
|
|
PwC-Embedding_expr was evaluated on the Korean subset of MTEB.
|
|
A leaderboard link will be added once it is published.
|
|
|
|
| Task | PwC-Embedding_expr |
|
|
|------------------|--------------------|
|
|
| KLUE-STS | 0.88 |
|
|
| KLUE-TC | 0.73 |
|
|
| Ko-StrategyQA | 0.80 |
|
|
| KorSTS | 0.84 |
|
|
| MIRACL-Reranking | 0.72 |
|
|
| MIRACL-Retrieval | 0.65 |
|
|
| **Average** | **0.77** |
|
|
|
|
|
|
## Model
|
|
- Base Model: [intfloat/multilingual-e5-large-instruct](https://huggingface.co/intfloat/multilingual-e5-large-instruct)
|
|
- Model Size: 0.56B
|
|
- Embedding Dimension: 1024
|
|
- Max Input Tokens: 514
|
|
|
|
|
|
## Requirements
|
|
It works with the dependencies included in the latest version of MTEB.
|
|
|
|
|
|
## Citation
|
|
|
|
TBD (technical report expected September 2025) |