You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

CC BY-NC-SA 4.0 — non-commercial research and educational use only.

⚠️ v6.0 available with stronger homophone defense: thc1006/cyberpuppy-v6-bilingual

thc1006/cyberpuppy-v6-pinyin-lora. v6 uses LoRA r=64 (vs v5's r=32) and gives HED-COLD +1.91pt, TC homo +0.14pt, COLD +0.19pt at α=0.60. v5.1.1 still holds slightly higher PCR-ToxiCN (0.7162 vs 0.7119); use v5 if PCR is your priority.

CyberPuppy v5 — Pinyin LoRA (LoRA-B) | Homophone Attack Defense

Phonetic-invariant toxicity detection — homophones produce identical input by construction, achieving 97.2% prediction invariance against homophone substitution attacks.

🛡️ Part of the CyberPuppy v5 dual-LoRA ensemble

When attackers replace toxic characters with same-pronunciation alternatives (「勾史」for「狗屎」), the pinyin branch sees identical input ("gou shi") → attack neutralized.

Why Pinyin?

Chinese homophone attacks exploit the gap between characters and pronunciation:

Original (toxic)	Homophone (looks benign)	Pinyin (identical)
狗屎 (dog shit)	勾史 (hook history)	gou shi
死掉 (die)	四調 (four tones)	si diao
你媽 (your mom)	密碼 (password)	ni ma / mi ma
白癡 (idiot)	白吃 (eat free)	bai chi

By converting text to toneless pinyin before classification, all homophones map to the same input string. This provides inherent robustness that no amount of data augmentation can match.

How It Works

Input: "勾史一個"  (looks benign — "hook history one")
  ↓ pypinyin conversion
Pinyin: "gou shi yi ge"  (= "狗屎一個" = "a piece of shit")
  ↓ LoRA-B classification
Output: TOXIC ✓

The text LoRA-A might be fooled by the benign-looking characters, but the pinyin LoRA-B always sees the true phonetic content.

Performance

Standalone vs Ensemble

Metric	Pinyin-only (this)	Text-only (LoRA-A)	Ensemble α=0.75
COLD F1_w	0.7979	0.8385	0.8336
PCR-ToxiCN F1	—	0.6613	0.6890
TC homo abs F1	~0.82	0.7930	0.8380
Homo invariance	97.2%	~79%	~93%

Why Both LoRAs Are Needed

Pinyin alone (this model): High invariance but lower semantic accuracy (0.7979 vs 0.8385)
Text alone (LoRA-A): Better semantics but vulnerable to homophone attacks
Ensemble: Best of both worlds — semantic accuracy + phonetic robustness

Usage

⚠️ This model must be used together with LoRA-A (text). See the full ensemble code on the companion model page.

Pinyin Conversion

from pypinyin import pinyin, Style
import re

_HAN = re.compile(r"[\u3400-\u4dbf\u4e00-\u9fff\uf900-\ufaff]")

def to_pinyin(text):
    """Convert Chinese text to toneless pinyin. Non-Han chars kept as-is."""
    syls = []
    for ch in text:
        if _HAN.match(ch):
            py = pinyin(ch, style=Style.NORMAL, errors="ignore")
            if py and py[0] and py[0][0]:
                syls.append(py[0][0])
        elif ch.strip():
            syls.append(ch)
    return " ".join(syls)

# Examples
assert to_pinyin("狗屎") == to_pinyin("勾史") == "gou shi"
assert to_pinyin("死掉") == to_pinyin("四調") == "si diao"

Training Details

Parameter	Value
Base model	Qwen/Qwen3-8B-Base
LoRA rank	32
LoRA alpha	64
Training data	179,186 samples (pinyin-converted)
Epochs	3
Learning rate	3e-5
Max length	128 tokens (pinyin is shorter)
Precision	bf16
Loss	Focal (γ=2.5) + uncertainty multi-task
Hardware	1× NVIDIA RTX 5090

Limitations

Lower standalone accuracy than text LoRA (0.7979 vs 0.8385) — designed for ensemble use, not standalone
Non-homophone attacks not addressed: Number substitution (4=死), letter substitution (X=逼), and creative slang don't produce different pinyin
Polyphones (多音字): Characters with multiple pronunciations may get wrong pinyin → wrong classification
Cantonese/Hokkien: Only covers Mandarin pinyin. Other Chinese dialect pronunciations not supported
Requires pypinyin dependency: Must convert text to pinyin at inference time (adds ~1ms latency)

Ethical Considerations

Same considerations as LoRA-A
The pinyin approach could theoretically be reversed to generate homophone attacks. Mitigated by non-commercial license.
Pinyin conversion is lossy (tone information removed) — model cannot distinguish tonal minimal pairs

Citation

@misc{cyberpuppy_v5_2026,
  author       = {Tsai, Hung-Che},
  title        = {CyberPuppy v5: Bilingual Dual-LoRA Ensemble for Chinese Cyberbullying Detection with Homophone Robustness},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/thc1006/cyberpuppy-v5-bilingual}}
}

Related Models

Model	Role
thc1006/cyberpuppy-v5-bilingual	Required companion text LoRA-A

Contact & Takedown

Author: Hung-Che Tsai (hctsai1006@cs.nctu.edu.tw)
Takedown: Email above — removed within 7 days

Downloads last month: -

Model tree for thc1006/cyberpuppy-v5-pinyin-lora

Base model

Qwen/Qwen3-8B-Base

Finetuned

(503)

this model

Dataset used to train thc1006/cyberpuppy-v5-pinyin-lora

Evaluation results

F1 (weighted, pinyin-only) on COLD
test set self-reported

0.798
Prediction Match Rate (homo pairs) on ToxiCloakCN (heldout)
test set self-reported

0.972
F1 (weighted, exceeds SOTA) on PCR-ToxiCN
self-reported

0.689