KLUE-NER SNRoberta

This project implements Named Entity Recognition for the KLUE (Korean Language Understanding Evaluation) benchmark using an ensemble of RoBERTa models. The project includes multi-GPU training automation and ensemble evaluation with Stochastic Weight Averaging (SWA).

Project Structure

.
├── README.md
├── environment.yaml
├── main_SWA_NER.py
├── eval_across_architecture.py
├── run_parallel_experiments-epochs=10_SWA_ensemble.sh
├── eval.sh
├── src/
│   ├── datasets.py
│   └── utils.py
└── data/
    └── klue-ner/

Installation

1. Clone the repository:

git clone https://github.com/jhpark-kaist/KLUE-NER-SNRoberta
cd KLUE-NER-SNRoberta

2. Install required packages:

conda env create -f environment.yaml
conda activate kllm-fairness

Usage

Training

chmod +x run_parallel_experiments-epochs=10_SWA_ensemble.sh
./run_parallel_experiments-epochs=10_SWA_ensemble.sh

Evaluation

chmod +x eval.sh
./eval.sh

Results

Ensemble Performance Across Seeds

Seed	Entity F1	Char F1
1	89.34	93.91
2	89.80	94.06
3	89.41	93.95
4	89.55	93.93
5	89.79	94.10
6	89.23	94.02
7	89.53	94.03
8	89.43	94.03

Average Performance Metrics

Metric	Score ± Variance
Entity F1	89.51 ± 0.19
Char F1	94.00 ± 0.06

Contact

Name : Joonhyeong Park
E-mail : jhpark.kaist@gmail.com
Affiliation : Statistical Inference and Machine Learning Lab @ KAIST