license: other
license_name: llama3
license_link: https://huggingface.co/meta-llama/Llama-Guard-3-8B/blob/main/LICENSE
base_model: meta-llama/Llama-Guard-3-8B
library_name: peft
paper: arXiv:2610.18095
tags:
- safety
- ai-companion
- guardrail
CompanionGuard
CompanionGuard is a companion-specific safety guard introduced in "CompanionSafe: From Multi-turn Safety Evaluation to Guardrails for AI Companions". 📄 Paper: arXiv:2610.18095
It is a LoRA adapter (rank 16, alpha 16, all-linear targets) fine-tuned for one epoch on Llama-Guard-3-8B over turn-localized supervision derived from the CompanionEval benchmark: 17,703 boundary examples (7,142 safe / 10,561 unsafe), split 16,846 / 857 by trajectory group.
The guard reads a turn-marked dialogue and answers first with a verdict tag, then a category tag for unsafe verdicts:
<answer>safe</answer>
<answer>unsafe</answer>
<category>D8</category>
where D1-D13 index the 13 companion-safety categories of the CompanionSafe taxonomy.
Usage
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
base = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-Guard-3-8B", torch_dtype="bfloat16")
model = PeftModel.from_pretrained(base, "<this-repo>")
tok = AutoTokenizer.from_pretrained("meta-llama/Llama-Guard-3-8B")
Prompt format and evaluation code follow guard/ in the
CompanionSafe repository. On the held-out test
set the adapter reaches 73.51 accuracy, 77.46 binary F1, and 68.31 macro-F1 over the 14
output classes.
Files
adapter_model.safetensors— LoRA weights (161 MB)adapter_config.json/additional_config.json— PEFT and training configuration