KoDEBATE-pre-v0.9-base — DEBATE-kor-initialized base model

This model was initialized from jongrock17/DEBATE-kor-base and further trained on the 71,216-pair training split of KoPolNLI-pre-v0.9.

The binary label mapping is 0 = not_entailment, 1 = entailment.

Provisional test results

Metric	Value
Test pairs	8,904
Accuracy	0.925764
Balanced accuracy	0.925764
Macro F1	0.925751
Entailment F1	0.924792
Non-entailment F1	0.926710
MCC	0.851812
AUROC	0.978466

Intended label	Predicted non-entailment	Predicted entailment
Non-entailment	4179	273
Entailment	388	4064

These are provisional scores against generator-intended labels, not human-verified ground truth. The test split was released before human review. Do not present these numbers as final benchmark performance.

한국어: Test 레이블은 생성 의도 레이블로, 인간검수 전입니다. 위 수치는 잠정 결과이며 확정 성능이 아닙니다.

Downloads last month: 24

Safetensors

Model size

0.2B params

Tensor type

F32

Model tree for jongrock17/KoDEBATE-pre-v0.9-base

Base model

mlburnham/Political_DEBATE_DeBERTa_base_v1.1

Finetuned

jongrock17/DEBATE-kor-base

Finetuned

(1)

this model

Dataset used to train jongrock17/KoDEBATE-pre-v0.9-base

Collection including jongrock17/KoDEBATE-pre-v0.9-base

KoPolNLI & KoDEBATE

Collection

Korean language and Korean politics specified NLI project (Dataset and zero-shot classifier) • 5 items • Updated 28 days ago