llama3.1-Instruct-qlora-cyberbullying

Llama 3.1 Instruct QLoRA – Cyberbullying Classifier

LoRA adapter for moderating social media posts. The model takes a post and returns a structured verdict in the format:

label: bully|not_bully; type: age|gender|ethnicity|religion|none

If bullying is detected, the adapter also predicts the bullying subtype; otherwise type: none is returned.

This adapter was fine-tuned from meta-llama/Meta-Llama-3.1-8B-Instruct using QLoRA with LLaMA-Factory. Training data is derived from Kaggle’s Cyberbullying Classification dataset.

TL;DR

• Task: binary bully/not_bully classification + bullying subtype (age, gender, ethnicity, religion)
• Training set: 35,883 posts
• Validation set: 3,986 posts
• Prompt: single system + instruction message with deterministic output template (see below)
• Result: 95.1% joint label/type accuracy on validation

Prompt template

System

You are a careful, consistent content‑moderation assistant. First decide whether the post is bullying; if it is, assign the most specific bullying type (age, gender, ethnicity, or religion). If it is not bullying, use type: none. Always output a single line exactly in this format: label: <label>; type: <type>, with no additional text or punctuation.

Instruction

Classify the following social media post. Return exactly one line in the format: label: bully|not_bully; type: <age|gender|ethnicity|religion|none>. Use lowercase exactly as shown. If the post is not bullying, set type to none. Do not include any extra words, explanations, or whitespace before or after the answer.

Training configuration

• Framework: LLaMA-Factory (QLoRA stage)
• Base model: Meta-Llama-3.1-8B-Instruct
• Adapter type: LoRA (rank 8, alpha 16, dropout 0.05) — DoRA disabled for QLoRA
• Quantization: 4-bit bnb
• Cutoff length: 2,048 tokens
• Optimizer: AdamW (fused)
• Learning rate: 5e-5, cosine scheduler, warmup ratio 0.05
• Batch size: 1 sample × 8 grad accumulation (effective 8)
• Epochs: 3 (seed 42)
• Precision: bfloat16

Validation loss reached 0.018 by the end of training (see trainer_state.json).

Evaluation results

Validation split: 3,986 posts (10% stratified sample, same prompt formatting).

Joint label+type accuracy is therefore 95.1%. The adapter maintains high recall for protected classes and is intentionally conservative on the none category (non-bullying content).

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = "meta-llama/Meta-Llama-3.1-8B-Instruct"
adapter = "muditbaid/llama3.1-Instruct-qlora-cyberbullying"

tokenizer = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(base, torch_dtype="auto")
model = PeftModel.from_pretrained(model, adapter)

system = ("You are a careful, consistent content-moderation assistant..." )  # see full text above
instruction = ("Classify the following social media post..." )
post = "Why are you acting like such a clown, grandma?"

prompt = (
    "<|begin_of_text|><|start_header_id|>system<|end_header_id|>\n\n" + system +
    "<|eot_id|><|start_header_id|>user<|end_header_id|>\n\n" + instruction + "\n" + post +
    "<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
)

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=16, do_sample=False)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

LLaMA-Factory usage

Add this repository as adapter_path in a LLaMA-Factory config (or pass --adapter_path) to evaluate or continue training.

Dataset notes

• Source: Kaggle Cyberbullying dataset (5 classes: age, gender, ethnicity, religion, not_cyberbullying)
• Pre-processing: normalization to label ∈ {bully, not_bully}; map not_cyberbullying → none subtype; remove rows with missing text
• Split: 90/10 stratified on class_label

Limitations

• English-only and trained on informal social media text
• Model may reproduce harmful language when prompted adversarially
• Designed for decision support; keep a human reviewer in the loop

Citation

@misc{muditbaid2024llama31cyberbullying,
  title  = {Llama 3.1 Instruct QLoRA – Cyberbullying Classifier},
  author = {Baid, Mudit},
  year   = {2024},
  howpublished = {Hugging Face Hub},
  url    = {https://huggingface.co/muditbaid/llama3.1-Instruct-qlora-cyberbullying}
}

---

Trained with PEFT 0.17.1, Transformers 4.57.1, PyTorch 2.9.0+cu128, Datasets 4.0.0, Tokenizers 0.22.1.