erikaecl/hansen-grooming-lora-gemma4-e4b

Hansen Grooming LoRA Adapter — Gemma 4 E4B + LoRA (4-bit)

LoRA adapter fine-tuned on top of google/gemma-4-E4B for binary classification of online grooming conversations.

Training Details

• Base model: google/gemma-4-E4B
• Method: QLoRA (4-bit NF4, rank=16, alpha=32)
• Task: Binary sequence classification (Safe vs Grooming)
• Dataset: PAN12 + AOL + NPS Chat + synthetic negatives (anonymized)

Usage

Gemma 4 has no native AutoModelForSequenceClassification. Load with lora_loader.py:

from lora_loader import load_lora_classifier
from grooming_lora_config import build_input

model, tokenizer, device = load_lora_classifier("erikaecl/hansen-grooming-lora-gemma4-e4b", device="cuda")
# model = Gemma4SequenceClassifier (LoRA backbone + classifier.bin head)

Adapter files: PEFT LoRA weights + classifier.bin + classifier_config.json.