tojohere/goemotions-ko-xlm-roberta-base

GoEmotions Korean — xlm-roberta-base (v2)

Fine-tuned xlm-roberta-base for 28-emotion classification on Korean text. Used in SentiLog for on-device journaling sentiment analysis.

Model Details

• Base model: FacebookAI/xlm-roberta-base (270M params)
• Training data: GoEmotions (EN+KO bilingual) + KOTE (50K native Korean emotions)
• Translation: NLLB-200-distilled-600M for EN→KO
• Training: 8 epochs with early stopping (patience=2, best at epoch 6)
• Quantization: INT8 dynamic quantization for ONNX (266MB)

Performance (v2)

Metric	v1	v2
F1 Macro (training)	0.379	0.568
28×3 Top-1	67%	68%
28×3 Top-3	83%	81%
Sanity Checks	8/8	8/8
Missed Emotions	2/28	2/28

Labels (28 GoEmotions)

admiration, amusement, anger, annoyance, approval, caring, confusion, curiosity, desire, disappointment, disapproval, disgust, embarrassment, excitement, fear, gratitude, grief, joy, love, nervousness, optimism, pride, realization, relief, remorse, sadness, surprise, neutral

Files

• model.safetensors — PyTorch checkpoint (for fine-tuning)
• goemotions.onnx — INT8 quantized ONNX (for inference, 266MB)
• tokenizer.json — Unigram tokenizer

Usage (ONNX)

import onnxruntime as ort
from tokenizers import Tokenizer

session = ort.InferenceSession("goemotions.onnx")
tokenizer = Tokenizer.from_file("tokenizer.json")

encoding = tokenizer.encode("오늘 정말 행복해요")
import numpy as np
input_ids = np.array([encoding.ids], dtype=np.int64)
attention_mask = np.array([encoding.attention_mask], dtype=np.int64)

outputs = session.run(None, {"input_ids": input_ids, "attention_mask": attention_mask})