Severity classifier for Nepali social media content, fine-tuned from XLM-RoBERTa Large using LoRA/PEFT. Part of the RakshakAI content moderation system.

📄 Paper: RakshakAI: Multi-Label Toxicity Detection for Low-Resource Nepali Social Media Content

What it does

Given a piece of Nepali text along with detected toxicity labels from rakshak-xlmr-large-v2, this model predicts one of three severity classes:

Class	Original Score	Meaning
`normal`	1–2	Clean or mildly offensive
`moderate`	3	Clearly harmful
`severe`	4–5	Incites action or extreme violence

How it differs from a standard classifier

Rather than predicting severity from text alone, the input is enriched by appending detected toxicity labels from the multi-label classifier to the original text — for example: yo jaat ko manchhe haru lai nepal bata nikala [DETECTED: hate_speech, casteism]

This allows the model to leverage category information when estimating severity, making predictions more accurate and context-aware.

Performance

Class	Precision	Recall	F1
Normal (1–2)	0.94	0.87	0.90
Moderate (3)	0.25	0.27	0.26
Severe (4–5)	0.86	0.91	0.88

The model performs strongly on normal and severe classes but struggles with moderate severity due to class imbalance — only 9.3% of training samples carry a moderate label. This is a known limitation we plan to address with more data collection.

How to use

python

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

model_id = "biraj-bhusal/rakshak-severity-v2"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSequenceClassification.from_pretrained(model_id)

id2label = {0: "normal", 1: "moderate", 2: "severe"}

# Append detected labels from rakshak-xlmr-large-v2 to the text
text = "your nepali text here [DETECTED: hate_speech, casteism]"
inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=128)

with torch.no_grad():
    logits = model(**inputs).logits
    prediction = torch.argmax(logits, dim=-1).item()

print(id2label[prediction])

Training Details

Base model: xlm-roberta-large
Fine-tuning method: LoRA (PEFT) — rank r=16, α=32, dropout=0.1
Training data: 4,716 augmented samples
Loss: Class-weighted cross-entropy to address class imbalance
Precision: fp16 mixed precision
Hardware: Dual NVIDIA T4 GPUs on Kaggle

Training Data

Limitations

Moderate severity class performs poorly (F1=0.26) due to underrepresentation in training data
Depends on output from rakshak-xlmr-large-v2 for best results
Single annotator labeling may introduce individual bias
May not generalize well to content outside Facebook and YouTube domains

Related Models

rakshak-xlmr-large-v2 — multi-label toxicity classifier whose output feeds into this model

Citation

bibtex

@misc{bhusal2025rakshak,
  author = {Bhusal, Biraj},
  title = {RakshakAI: Multi-Label Toxicity Detection for Low-Resource Nepali Social Media Content},
  year = {2025},
  publisher = {Zenodo},
  url = {https://zenodo.org/records/20850923}
}

Content Warning

This model is trained on toxic social media content. The author does not endorse any views expressed in the training data.

Developed by: Biraj Bhusal