Anthropic updated its usage policy to prohibit sustained and needless abusive or cruel behavior toward Claude. The new policy targets only extreme cases where users repeatedly act cruelly toward Claude with no discernible purpose.
The details:
Terminating conversations remains the primary enforcement mechanism for the new policy.
Anthropic consolidated existing restrictions into a new ban on deceptive commercial or political campaigns.
The updated policy expands the weapons ban to include software and components that make weapons work, including arming drones.
🎛️ Researchers propose risk-conditioned RLHF for LLMs
Mask Fine-Tuning (MFT) is a novel LLM fine-tuning paradigm introduced in the article. Large Language Models (LLMs) are increasingly deployed in settings where rare but severe harmful generations can have significant consequences.
The details:
Existing Risk-Averse RLHF addresses the issue of harmful generations by optimizing Conditional Value-at-Risk (CVaR).
Risk-Averse RLHF trains policies for fixed risk levels.
Risk-Averse RLHF cannot adjust the desired degree of risk aversion at inference time.
