erikaecl/hansen-grooming-lora

Hansen Grooming LoRA Adapter

LoRA adapter fine-tuned on top of Meta-Llama-3-8B for binary classification of online grooming conversations.

Training Details

• Base model: meta-llama/Meta-Llama-3-8B
• Method: LoRA (rank=16, alpha=32)
• Task: Binary sequence classification (Safe vs Grooming)
• Dataset: PAN12 + NPS Chat + synthetic negatives (anonymized)
• Training: 3 epochs, bf16, gradient checkpointing, weighted loss

Usage

import torch
from transformers import AutoModelForSequenceClassification, AutoTokenizer
from peft import PeftModel

base_model_id = "meta-llama/Meta-Llama-3-8B"
adapter_id = "erikaecl/hansen-grooming-lora"

tokenizer = AutoTokenizer.from_pretrained(base_model_id)
base_model = AutoModelForSequenceClassification.from_pretrained(
    base_model_id,
    num_labels=2,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
model = PeftModel.from_pretrained(base_model, adapter_id)
model.eval()

Dataset

Trained on anonymized data — all user handles replaced with generic labels (user_a, user_b, etc.) before training. No PII in the training set.