kmseong/llama3.2_3b_SSFT_epoch3_lr2e-5

Model Overview

The kmseong/llama3.2_3b_SSFT_epoch3_lr2e-5 is a 3.2 billion parameter model built upon the meta-llama/Llama-3.2-3B-Instruct base. It represents Phase 0 of the Safety-WaRP (Weight space Rotation Process), a specialized training methodology focused on enhancing model safety.

Key Characteristics

• Safety-Focused Training: The model has been fine-tuned using the Circuit Breakers dataset over 3 epochs to establish core safety mechanisms, aiming to produce refusal responses to harmful prompts.
• Llama 3.2 Architecture: Utilizes the Llama 3.2 architecture with bfloat16 precision and gradient checkpointing enabled.
• Foundational Safety Model: This version is intended as a base model with robust safety capabilities, designed to be a prerequisite for subsequent training phases that restore or enhance utility.

Important Considerations

• Utility Trade-off: As a Phase 0 model, its primary focus is safety. Consequently, its utility in areas like mathematics or reasoning may be reduced compared to general-purpose models.
• WaRP Pipeline: This model is part of a multi-phase WaRP pipeline. Future phases (Phase 1: Basis Construction, Phase 2: Importance Scoring, Phase 3: Incremental Learning) are designed to restore utility while maintaining safety.
• Usage: Developers can use this model for applications requiring a strong safety baseline, particularly as a component in a larger system where utility will be addressed in later stages.