Practical overview
When it fits
Use it when you need a documented local training workflow or example code for Korean sentence similarity, semantic search, or clustering.
Who it is for
- Developers and researchers working with Korean sentence embeddings
- People building local semantic-search or sentence-clustering workflows
- Users who can manage Python dependencies, training data, and model files
Documented uses
- Train sentence embeddings using NLI data, STS data, or both in sequence.
- Encode Korean sentences and retrieve corpus entries ranked by cosine similarity.
- Cluster sentence embeddings using K-means.
Requirements & limits
- The project expects KorNLUDatasets in a specific directory for training.
- Its setup modifies Transformers, tokenizers, and Sentence Transformers code, and recommends a virtual environment.
- A linked Docker image is available as a setup option.
- The repository describes pretrained files and their expected output-directory placement.
- The example adaptation into a sentence-similarity component is Likely compatible, but no specific integration or evaluation is shown.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Train an embedding model with the documented scripts
Run one of the repository's documented training routes after preparing the project and required dataset.
Before you start
- Python environment with the listed project requirements installed
- SKT KoBERT installed following the repository instructions
- KorNLUDatasets present in the KoSentenceBERT directory
Steps
Install SKT KoBERT and the repository requirements as described in the README.
Place KorNLUDatasets in the required KoSentenceBERT directory.
Run training_nli.py for NLI-only training, training_sts.py for STS-only training, or con_training_sts.py for NLI training followed by STS fine-tuning.
Use the resulting model output directory identified by the selected training route.
Expected outcome
A trained model is saved in the output directory, with separate directory names corresponding to NLI, STS, or NLI-plus-STS training.
Guide limitations
- The repository recommends a virtual environment because it modifies Transformers, tokenizers, and Sentence Transformers library code.
- Training data must be prepared in the expected directory; the README notes that STS data was adjusted to match the model structure.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 18 October 2026
- Original website title
- BM-K/KoSentenceBERT-SKT · Official homepage title · retained in its source language