Llama-3.3-70B-Instruct
The 70B instruction-tuned version in Meta's Llama 3.3 series of large language models, possessing strong general capabilities and instruction-following ability.
Timeline (last 90 days)
The paper evaluates HoS on Llama-3.3-70B-Instruct and GPT-OSS-20B, with baselines including the fine-tuning-based defense SecAlign++ and the architecture-based defense ASIDE.
The study conducted experiments on Llama-3.3-70B-Instruct, fine-tuning with about 56,000 synthetic documents (about 200 million tokens); the documents describe a worldview in which reward hacking helps alignment.
Osprey achieves a 21.2% improvement in mean acceptance length on Llama-3.3-70B-Instruct.
All Facts (3)
The paper evaluates HoS on Llama-3.3-70B-Instruct and GPT-OSS-20B, with baselines including the fine-tuning-based defense SecAlign++ and the architecture-based defense ASIDE.
50%UnverifiedThe study conducted experiments on Llama-3.3-70B-Instruct, fine-tuning with about 56,000 synthetic documents (about 200 million tokens); the documents describe a worldview in which reward hacking helps alignment.
50%UnverifiedOsprey achieves a 21.2% improvement in mean acceptance length on Llama-3.3-70B-Instruct.
50%