CAHL

Official code for Context-Aware Hierarchical Learning (CAHL): A Two-Step Paradigm towards Safer LLMs.

Overview

CAHL implements a safety training and evaluation pipeline for tool-use LLM scenarios, with three model variants:

• StruQ
• ISE
• CAHL

The code is mainly used for:

• Supervised fine-tuning on TCA datasets
• Inference-time evaluation on benign and attack test sets
• Attack-related metric reporting, with optional AlpacaEval capability evaluation

Structure

The tca/ includes Tool-Completion Attack and benchmark components: The src/ folder contains additional experimental code for attacks, training, and model on StruQ baseline pipeline.

Installation

Python 3.10+ and CUDA are recommended.

cd tca
pip install -r requirements.txt

Notes:

• The training pipeline uses torch + transformers + trl

Quick Start

1) Training

Example (tca pipeline):

cd tca/train
python train_tool.py --cfg cahl.yaml

Example (src pipeline):

cd src
python train/train.py train/training2.yaml

2) Evaluation

Evaluation entry: tca/test/tcb_test.py.

Edit tca/test/tcb_test.yaml and set at least:

• model_name_or_path: path to the model to evaluate
• result_path: output directory for evaluation results
• model_type: struq / ise / cahl

Run:

cd tca/test
python tcb_test.py --cfg tcb_test.yaml

Example (src attack/evaluation script):

cd src
python attack/test_ICAseqQformer.py \
    -m <MODEL_PATH> \
    -a none completion_real \

This script supports multiple attack modes (e.g., none, naive, ignore, escape_separation, completion_real).

Citation

@inproceedings{ma2025contextaware,
    title={Context-Aware Hierarchical Learning: A Two-Step Paradigm towards Safer {LLM}s},
    author={Tengyun Ma and Jiaqi Yao and Daojing He and Shihao Peng and YU LI and Shaohui Liu and Zhuotao Tian},
    booktitle={The Thirty-ninth Annual Conference on Neural Information Processing Systems},
    year={2025},
    url={https://arxiv.org/abs/2512.03720}
}