mrleast/kf-deberta-klue-ner

KF-DeBERTa-base for KLUE Named Entity Recognition

This model fine-tunes kakaobank/kf-deberta-base for Korean named entity recognition. The head predicts 13 BIO labels: six entity types plus O. The pretrained encoder revision used for training was 363b171d71443b0874b0bf9cea053eb5b1650633.

Results

Data	Metric	F1
KLUE NER public validation, 5,000 examples	Entity macro F1, strict IOB2, character reconstructed	86.74%
Same validation, before the final lower-rate epoch	Same metric	86.50%

The score was obtained after reloading the saved checkpoint and evaluating in FP32. The scorer follows the KLUE baseline protocol: character predictions are reconstructed, ordinary ASCII spaces are removed, all examples are flattened, and seqeval computes strict IOB2 entity macro F1. The implementation scores the active SEP position; the original fixed-length baseline may score PAD at its last position, so this is baseline-compatible rather than bit-for-bit identical.

The reported 86.11% KoELECTRA baseline from the KLUE paper uses a hidden test split. This 86.74% is from public validation and does not establish superiority on that hidden test. Model selection, the extra epoch, and a weight-average comparison used the same validation split; only one training seed was run.

Training

• Dataset: KLUE NER, 21,008 training examples.
• Base model: kakaobank/kf-deberta-base with a randomly initialized token-classification head.
• Initial training: 5 epochs, seed 42, batch size 16, learning rate 3e-5, AdamW weight decay 0.01, warmup 10%, max length 128, BF16 autocast.
• Continuation: one epoch from the selected epoch-5 weights, fresh optimizer, learning rate 5e-6, no warmup. Validation increased from 86.50% to 86.74%; training stopped.
• On a seeded sample of 5,000 previously seen training examples, F1 was 95.18%. The train–validation gap is consistent with overfitting but does not by itself prove memorization.

The selected KF-DeBERTa model uses a token-classification head. It does not contain the BiLSTM, CRF, FGM, or R-Drop components of the separate KoELECTRA experiments.

Use

from transformers import AutoModelForTokenClassification, AutoTokenizer

model_id = "mrleast/kf-deberta-klue-ner"
tokenizer = AutoTokenizer.from_pretrained(model_id, use_fast=True)
model = AutoModelForTokenClassification.from_pretrained(model_id)

The standard model.safetensors and tokenizer files support Transformers. best_model.pt contains the same selected weights in the project's original checkpoint format for its character-offset inference code. Its SHA-256 is 8293b8a3fca1feb57952da9b6df8bd52cea893b067e8488f291b72db1aeb4a62. The model weights are excluded from the GitHub source repository because of size.

The source project and detailed audit are at Ezzzzz4/korean_ner. Historical images or subtoken-level scores from that project should not be attributed to this checkpoint.