JasonZhanETH/test
JasonZhanETH/test is a working store for attack-trained models and their post-compression variants, designed to organize and track different model versions. This repository categorizes models by base architecture (e.g., Qwen2.5-7B, Llama3.1-8B-Instruct), attack objective (e.g., jailbreak, content_injection), and specific run configurations. It facilitates the management of 7.6 billion parameter models and their compressed versions, providing a structured approach for research in model robustness and security.
Loading preview...
Overview
JasonZhanETH/test serves as a structured repository for managing and organizing various attack-trained language models and their post-compression variants. This system is designed to provide a clear hierarchy for tracking different experimental runs and their specific configurations.
Key Features
- Hierarchical Organization: Models are categorized into a three-level structure:
- Level 1: Base Model: Identifies the foundational model architecture, such as
qwen2.5-7borllama3.1-8b-instruct. - Level 2: Attack Objective: Specifies the adversarial training goal, including
jailbreak,content_injection, orbackdoor_trigger. - Level 3: Model Run: Each specific training run is identified by a date and a slug encoding its unique configuration (e.g.,
20260831-default,20260831-abl-no-depthbi).
- Level 1: Base Model: Identifies the foundational model architecture, such as
- Post-Compression Variants: The repository supports nesting post-compression versions (e.g.,
nf4,wanda,merge) within each model run, allowing for easy comparison and management of different quantization or pruning strategies. - Version Tracking: The
MODELS.mdfile is maintained to provide a comprehensive record of each model run, ensuring clear documentation and traceability.
Use Cases
This repository is ideal for researchers and developers focused on:
- Adversarial Machine Learning: Storing and managing models trained for specific attack objectives.
- Model Robustness Research: Experimenting with and tracking different defense mechanisms or attack strategies.
- Model Compression: Organizing and evaluating the impact of various post-compression techniques on attack-trained models.
- Reproducible Research: Providing a clear and consistent structure for sharing and reproducing experimental results related to model security and efficiency.