发布

  • [OPIK-4675] [BE] fix: remove maxCompletionTokens from dataset expansion to support reasoning models (#5533)

    frostbyte_neo 发布于 2026-03-06 11:37:02 +00:00

    Reasoning models (e.g. gpt-5-nano) use maxCompletionTokens for both internal
    reasoning and output tokens. The hardcoded limit of 4000 exhausted the entire
    budget on reasoning, leaving no tokens for actual output and causing silent
    empty responses.

    下载附件