Open Questions on gpt-oss-safeguard implementation

[andrewmchang]

Starting a thread for questions on gpt-oss-safeguard implementation!

• (Overheard) Is the harmony format applied to the message content? The OpenAI Go library supports the roles already, so not sure where to put the harmony format
• (Overheard) How can you improve latency of gpt-oss-safeguard responses? --> Ensure prompts/policies are the right length, as detailed in the user guide (i.e., 400-600 tokens per policy) + ensure use of harmony format

[andrewmchang]

Re: latency, one comment from Kosta: "It turns out getting the policy down to 400 tokens from 900 tokens made a significant improvement, and switching to reasoning effort low (instead of medium) also improved latency without any decrease in accuracy (so far, anyway)"

[andrewmchang]

X-posting from the RMC Discord: "Has anyone had success in getting gpt-oss-safeguard to understand the current date? In flagging content, one of its justifications is often "This is an impossible future-dated event". What I've been doing is passing the current timestamp to the policy, and the policy has The current date and time is provided in metadata.current_date. But I am also having trouble getting gpt-oss-safeguard to consider the fields provided in the metadata."