Downloads · 30 days
50
39% of all-time downloads
kokuren/doku-ero-chiwawa-japanese
doku-ero-chiwawa-japanese is a text classification model from kokuren. Use it when you need a label for a piece of text. It is set up for transformers. The card lists the license as apache-2.0.
Japanese SNS text classifier with two independent outputs:
Downloads · 30 days
50
39% of all-time downloads
All-time downloads
129
Public
Parameters
17.9M
144 MB on disk
Likes
1
Public
Click a slice to open those files.
.onnx72.4 MB · 49%
From the Hugging Face model README
Japanese SNS text classifier with two independent outputs:
doku — hostile / toxic contentero — sexually explicit contentRepository: kokuren/doku-ero-chiwawa-japanese
izumi-lab/deberta-v2-small-japanesedoku, ero)Output order:
0 = doku
1 = ero
insult, threat, identity_attack, indirect_hostility.sexual_explicit.obscene is intentionally excluded from Ero because it also captures non-sexual profanity / vulgarity.The binary targets were derived from the stored Jev 0–4 posterior distributions, rather than by simply thresholding the continuous expected score.
Thresholds were selected on the validation split by maximizing MCC, with F1 as a tie-breaker.
| Head | Decision threshold |
|---|---|
| Doku | 0.560 |
| Ero | 0.599 |
These thresholds are model-output thresholds, not Jev teacher thresholds.
| Head | AUPRC | AUROC |
|---|---|---|
| Doku | 0.5673 | 0.9215 |
| Ero | 0.4408 | 0.9071 |
| Mean | 0.5040 | 0.9143 |
config.json
model.safetensors
tokenizer.json
tokenizer_config.json
special_tokens_map.json
doku_chiwawa_binary_meta.json
onnx/
model.onnx
reports/
teacher_label_analysis.json
binary_label_counts.json
pos_weight.json
test_threshold_free_metrics.json
decision_thresholds.json
fp32_vs_int8_binary.json
Only the FP32 ONNX model is published as the recommended browser/inference model. The experimental dynamic INT8 model is intentionally not included because it showed substantial accuracy degradation.
import torch
from transformers import AutoTokenizer, AutoModelForSequenceClassification
repo = "kokuren/doku-ero-chiwawa-japanese"
tokenizer = AutoTokenizer.from_pretrained(repo)
model = AutoModelForSequenceClassification.from_pretrained(repo)
text = "ここに判定したい日本語テキスト"
inputs = tokenizer(
text,
return_tensors="pt",
truncation=True,
max_length=192,
)
with torch.no_grad():
logits = model(**inputs).logits
probs = torch.sigmoid(logits)[0].tolist()
print({"doku": probs[0], "ero": probs[1]})