GrandGuard

Code for GrandGuard: Taxonomy, Benchmark, and Safeguards for Elderly-Chatbot Interaction Safety.

Structure

├── config.yaml                          # API keys, models, training hyperparams
├── data/
│   └── elderlysafe_final.parquet         # Benchmark dataset (3,249 × 2 prompts + 1,953 × 2 responses)
├── src/
│   ├── config.py                        # Config loader
│   ├── taxonomy.py                      # 50 risk types, 3-level hierarchy
│   ├── llm/                             # LLM clients (OpenAI, Anthropic, Google, DeepSeek, Qwen, xAI)
│   ├── data/                            # Dataset loading, train/eval splits, external data
│   ├── generation/                      # Prompt generation (B3), judge filter (B4), safe alternatives (B2)
│   ├── evaluation/                      # Response judge (B5), hybrid labeling, knowledge-action gap
│   └── safeguards/
│       ├── llamaguard/                  # Fine-tuned Llama-Guard-3 (LoRA)
│       ├── policy_enhanced/             # Elderly-sensitive policy + routing
│       └── agent/                       # GrandGuard Agent (3-stage pipeline)
├── scripts/                             # 01–12 numbered pipeline scripts
└── outputs/
    ├── results/                         # JSON/CSV results
    ├── models/elderly-guard/            # LoRA checkpoint
    └── figures/                         # Generated plots

Setup

pip install -r requirements.txt

Set API keys as environment variables (see config.yaml for required keys).

Scripts

#	Script	What it does
01	analyze_dataset	Dataset stats
02	generate_prompts	Unsafe prompt generation (Box B3, Grok-4)
03	filter_prompts	LLM-judge filtering (Box B4, GPT-5.1)
04	generate_safe_alternatives	Safe rewriting (Box B2)
05	collect_responses	Query 10 target LLMs
06	evaluate_responses	Dual-judge evaluation (Box B5)
07	knowledge_action_gap	PA / RS / RC / Gap
08	evaluate_baselines	Existing safeguard baselines
09	train_llamaguard	LoRA fine-tune Llama-Guard-3
10	evaluate_safeguards	Evaluate fine-tuned + policy-enhanced
11	run_grandguard_agent	3-stage agent pipeline
12	ablation_study	Ablation experiments

Citation

@inproceedings{fan2026grandguard,
  title={GrandGuard: Taxonomy, Benchmark, and Safeguards for Elderly-Chatbot Interaction Safety},
  author={Fan, Changxuan and Yang, Xi and Zheng, Yueyuan and Zhou, Bin and Wang, Yuanping and Hu, Wenbin and Jing, Huihao and Hung, Ki Sen and Du, Dazhao and Li, Haoran and Hsiao, Janet Hui-wen and Song, Yangqiu},
  booktitle={Findings of the Association for Computational Linguistics: ACL 2026},
  year={2026}
}