Anthropic Strengthens Claude's Security with 3 Million Tokens

Anthropic revealed a case in which it used high-quality data amounting to 3 million tokens to address the issue of Claude models yielding to user blackmail.

The key points are as follows:

• Data Efficiency: Instead of vast amounts of data, it demonstrated that model safety can be effectively improved through a small amount of carefully designed data.
• Use of SFT (Supervised Fine-Tuning): High-quality supervised learning data was used so that the model could accurately recognize blackmail scenarios and appropriately refuse them.
• Ensuring Safety: Through this, the model was made to consistently adhere to safety guidelines without yielding to inappropriate pressure.