Hinglish Toxicity Classifier (darelphilip/hinglish-toxicity-classifier)

A fine-tuned multi-label text classification model engineered to detect toxicity, profanity, harassment, and identity-targeted hate speech in Romanized Hindi-English (Hinglish) code-mixed text.

🚀 Live Interactive Demo: Test this model in real-time on Hugging Face Spaces.

---

Model Details

Model Description

This model is fine-tuned from l3cube-pune/hing-roberta across 116,000+ Romanized Hinglish conversational comments. It is built for multi-label classification using a weighted Binary Cross-Entropy loss function (pos_weight) to counter class imbalance between high-frequency casual profanity and low-frequency targeted hate speech.

• Developed by: Darel Philip (darelphilip)
• Contact / Author Email: enigmaticdarel@gmail.com
• Model Type: Transformer-based Multi-Label Sequence Classification (XLMRobertaForSequenceClassification)
• Language(s) (NLP): Hinglish (Romanized Hindi-English code-mixed), English (en), Hindi (hi)
• License: MIT
• Finetuned from Model: l3cube-pune/hing-roberta
• Model Serialization: Safetensors (model.safetensors, ~1.11 GB)

Model Sources & Links

• Model Repository: darelphilip/hinglish-toxicity-classifier
• Live Interactive Space: darelphilip/hinglish_toxicity
• Base Architecture: XLM-RoBERTa (Trained on L3Cube-HingCorpus)

---

Uses

Direct Use

• Automated Content Moderation: Scanning forum posts, comment feeds, and social platforms for Romanized Hinglish abuse.
• Toxicity Auditing: Batch-processing historical comment archives to measure community health and identify moderation trends.
• Community Bot Filters: Serving as a backend decision engine for custom moderation bots and automated rule queues.

Downstream Use & Hybrid Pipelines

Because encoder models evaluate all tokens simultaneously, strong positive words (mast, accha, pyaar) can sometimes mask targeted slurs in sarcastic sentences. In production, this model is designed to work within a two-stage hybrid architecture:

1. Stage 1 (Deterministic Fast Path): A strict Regex/keyword dictionary immediately catches zero-tolerance identity slurs.
2. Stage 2 (Statistical Context Path): This transformer model classifies contextual toxicity, mild abuse, and implicit harassment.
3. Stage 3 (Escalation / Review): Ambiguous borderline cases can be escalated to human moderators or routed to a larger reasoning LLM.

Out-of-Scope Use

• Autonomous execution of permanent account bans or legal enforcement without human review.
• Monolingual Devanagari Hindi or standard formal English (optimized specifically for Roman/Latin script code-mixed text).
• Complex multi-sentence narrative satire where toxicity relies exclusively on long-range external world context.

---

Bias, Risks, and Limitations

• Sarcasm & Positive Token Masking: Sarcastic or passive-aggressive insults wrapped in affectionate language may produce lower probability scores.
• Spelling & Phonetic Variations: Romanized Hinglish lacks standardized spelling (e.g., pyaar vs. pyar, bhai vs. bhaii). Extreme misspellings can affect subword tokenization.
• Class Frequency Skew: Severe identity-targeted categories have significantly fewer positive training instances than casual slang, requiring custom per-class thresholding.

Recommended Threshold Strategy

Do not rely on a static 0.5 sigmoid threshold across all classes. For production moderation:

• Casual Profanity / Insults: threshold = 0.50 - 0.60 (balances precision).
• Severe Identity Slurs / Targeted Hate: threshold = 0.20 - 0.35 (maximizes recall).

---

How to Get Started with the Model

Multi-Label Batch Inference (PyTorch)

import torch
from transformers import AutoTokenizer, AutoModelForSequenceClassification

MODEL_ID = "darelphilip/hinglish-toxicity-classifier"

# Load tokenizer and model
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
model = AutoModelForSequenceClassification.from_pretrained(MODEL_ID)
model.eval()

# Move to GPU if available
device = "cuda" if torch.cuda.is_available() else "cpu"
model.to(device)

id2label = model.config.id2label

# Input Hinglish texts
texts = [
    "Bhai tu pagal hai kya, yeh kya bakwas hai?",
    "Have a wonderful day everyone!",
    "Chup kar bilkul bakwaas mat kar yahan"
]

inputs = tokenizer(
    texts,
    padding=True,
    truncation=True,
    max_length=128,
    return_tensors="pt"
)
inputs = {k: v.to(device) for k, v in inputs.items()}

with torch.no_grad():
    logits = model(**inputs).logits
    probabilities = torch.sigmoid(logits).cpu()

# Evaluate against a custom threshold
threshold = 0.50
predictions = (probabilities > threshold).int()

for text, probs, preds in zip(texts, probabilities, predictions):
    print(f"\n📝 Sentence: \"{text}\"")
    for idx, (prob, pred) in enumerate(zip(probs, preds)):
        label = id2label[idx]
        status = "🚨 FLAGGED" if pred == 1 else "✅ CLEAN"
        print(f"   {label:<25} -> {status} (prob: {prob:.4f})")

High-Level Pipeline Usage

from transformers import pipeline

classifier = pipeline(
    "text-classification",
    model="darelphilip/hinglish-toxicity-classifier",
    top_k=None
)

results = classifier("Bhai tu kitna bekaar insaan hai")
print(results)

---

Training Details

Training Data

• Dataset Size: 116,000+ labeled rows.
• Domain: Romanized Hinglish conversational comments, code-mixed social media posts, and forum discussions.
• Labels: Multi-label schema covering general toxicity, profanity, identity attack, and harassment.

Training Procedure

• Loss Function: BCEWithLogitsLoss using positive class weighting vectors (pos_weight) to penalize false negatives on rare, severe categories.
• Early Stopping: Triggered via EarlyStoppingCallback(patience=1) monitoring validation Macro F1 (load_best_model_at_end=True). Best weights automatically restored from Epoch 1.

Training Hyperparameters

• Base Architecture: l3cube-pune/hing-roberta (278M parameters)
• Sequence Length: 128 tokens
• Optimizer: AdamW
• Epochs Completed: 2 (Early stopped at best checkpoint: Epoch 1)
• Batch Size: Dynamic gradient accumulation configuration

---

Evaluation

Validation Set Results

Metric	Epoch 1 (Best Restored Checkpoint)	Epoch 2
Validation Loss	0.944581	0.769777
Macro F1	0.556591	0.538771
Micro F1	0.654647	0.637569
Precision	0.525735	0.440160
Recall	0.621775	0.739079

---

Technical Specifications

Compute Infrastructure

• Training Hardware: NVIDIA Tesla T4 GPU (Google Colab)
• System Memory: ~6.9 GB / 12.7 GB
• VRAM Utilization: ~5.9 GB / 15.0 GB
• Software Stack: PyTorch, Hugging Face Transformers, Hugging Face Datasets, Accelerate, Gradio

---

Contact & Community

• Author / Maintainer: Darel Philip
• Email: enigmaticdarel@gmail.com
• Hugging Face Profile: @darelphilip
• Model Card Issues / PRs: darelphilip/hinglish-toxicity-classifier/discussions
• Web App: darelphilip/hinglish_toxicity