Overview of TARS-SFT-1.5B

TARS-SFT-1.5B is a 1.5 billion parameter language model specifically designed for safety-oriented reasoning tasks. It functions as the (\pi_{SFT}) component within the TARS (Training Adaptive Reasoners for Safety) framework, which is detailed in the research paper "Reasoning as an Adaptive Defense for Safety" (arXiv:2507.00971).

Key Capabilities & Features

• Safety-tuned Reasoning: This model is SFT-tuned (Supervised Fine-Tuning) with a focus on improving reasoning abilities for safety applications.
• RL Base Model: It serves as the foundational model for subsequent Reinforcement Learning (RL) training within the TARS methodology.
• Origin: Developed by danielkty22, it is built upon the robust architecture of Qwen2.5-1.5B-Instruct.
• High Context Length: Features a substantial context window of 131,072 tokens, enabling processing of extensive inputs.

Ideal Use Cases

• Research in AI Safety: Particularly suited for researchers exploring adaptive defenses and reasoning in AI safety.
• Foundation for RL Training: Excellent as a starting point for further RL fine-tuning to develop more sophisticated safety reasoners.
• Applications Requiring Robust Reasoning: Useful in scenarios where a lightweight model needs strong reasoning capabilities, especially in safety-critical contexts.