6 comments

  • aesthesia 2 hours ago
    My guess is that RL training being done with particular generation parameters makes models much more brittle to changes in these parameters, and that's why we're seeing changes like this across model providers. But I don't really know.
    • pixelmelt 2 hours ago
      I'm inclined to agree given how unstable Gemma 4 is when not using the "official" sampler settings
  • tolugenius 2 hours ago
    > To improve determinism, define a system instruction with explicit rules for your specific use case.

    Is this guaranteed to work any better than top_k or top_p? This just sounds like making a smaller version of a Agent.md doc.

    • janalsncm 21 minutes ago
      It is guaranteed to work worse than top_k=1, that’s for sure.
    • sara_mo 15 minutes ago
      [flagged]
  • kouteiheika 59 minutes ago
    Obligatory "The Conspiracy Against High Temperature Sampling":

    https://gist.github.com/Hellisotherpeople/71ba712f9f899adcb0...

    • NooneAtAll3 24 minutes ago
      where can one learn what top_k and top_p mean?
  • tough 2 hours ago
    fwiw sonnet-5 also drops temperature (sonne-4 had it)
  • impulser_ 2 hours ago
    Good. These have been basically useless for the past few generations of models, and most of the time made the model perform worst.
  • greatgib 4 hours ago
    1. Sampling parameter deprecation (temperature, top_p, top_k)

    temperature, top_p, and top_k are deprecated and ignored. In future model generations, supplying these parameters returns an HTTP 400 error. Remove these parameters from all requests.