Japanese X Ragebait - Tohoku BERT Base v3

This is the Tohoku BERT component of the three-model ragebait detector from “From Detection to Characterization: A Large-Scale Study of Ragebait on Japanese X” (WI-IAT 2026). It is a binary Japanese text classifier fine-tuned from tohoku-nlp/bert-base-japanese-v3.

Labels

• NO (0): non-ragebait
• YES (1): ragebait

The labels are research annotations and should not be treated as facts about a post or its author.

Training

• Labeled split: 16,558 posts, balanced between YES and NO
• Train/validation partition: 14,902 / 1,656
• Held-out test split: 2,000 posts, balanced between YES and NO
• Epochs: 3
• Learning rate: 3e-5
• Batch size: 32
• Maximum sequence length: 256
• Warmup ratio: 0.1
• Weight decay: 0.01
• Seed: 42
• Best checkpoint criterion: validation Macro-F1

The public dataset repository contains only Post IDs and labels, not redistributed post text: ZhiyangQi/japanese-x-ragebait.

Test results

The confusion matrix is [[798, 202], [130, 870]], ordered as NO, YES. Full training history and unrounded values are in training_results.json.

Usage

import torch
from transformers import AutoModelForSequenceClassification, AutoTokenizer

model_id = "ZhiyangQi97/japanese-x-ragebait-tohoku-bert-base-v3"
tokenizer = AutoTokenizer.from_pretrained(model_id, use_fast=False)
model = AutoModelForSequenceClassification.from_pretrained(model_id)

text = "分類したい日本語テキスト"
inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=256)
with torch.inference_mode():
    prediction = model(**inputs).logits.argmax(dim=-1).item()

print(model.config.id2label[prediction])

Install transformers, torch, fugashi, and unidic-lite before loading the tokenizer.

Limitations and intended use

This model is intended for non-commercial academic research on Japanese ragebait. The training labels were created with an LLM-assisted pipeline and may contain errors. Ragebait is contextual and intent-based; performance may degrade on other periods, platforms, languages, or domains. Do not use this model to profile, rank, target, identify, harass, moderate, or make consequential decisions about individual users.

License

The base model is distributed under the Apache License 2.0. See the base model card for its terms and documentation.

Citation

@inproceedings{qi2026ragebait,
  title = {From Detection to Characterization: A Large-Scale Study of Ragebait on Japanese X},
  author = {Qi, Zhiyang and Ito, Kazuhiro and Chen, Jinghui and Nakamura, Hibiki and Chen, Zhangxuan and Murata, Erina and Chujyo, Masaki and Toriumi, Fujio},
  booktitle = {WI-IAT 2026},
  year = {2026}
}