- They might be dynamically adjusting these at inference time [1]. For example, start with a low temperature and generate samples with increasingly high temperatures until one of them passes some quality gate.
- They don't want you to fine-tune on high temperature completions (rejection fine-tuning).
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.
ironically, any frontier LLM will easily generate a tutorial at any detail you like explaining what these are.
if don't have time for that, just know that these are technical parameters that affect how likely it is an llm will produce the same result after being asked the same question.
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.
Possible reasons:
- They might be dynamically adjusting these at inference time [1]. For example, start with a low temperature and generate samples with increasingly high temperatures until one of them passes some quality gate.
- They don't want you to fine-tune on high temperature completions (rejection fine-tuning).
[1] https://rlhfbook.com/c/09-rejection-sampling#related-best-of...
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.
I'm inclined to agree given how unstable Gemma 4 is when not using the "official" sampler settings
Along with everything else. These parameters can make speculative decoding less accurate increasing the inference cost.
Obligatory "The Conspiracy Against High Temperature Sampling":
https://gist.github.com/Hellisotherpeople/71ba712f9f899adcb0...
where can one learn what top_k and top_p mean?
ironically, any frontier LLM will easily generate a tutorial at any detail you like explaining what these are.
if don't have time for that, just know that these are technical parameters that affect how likely it is an llm will produce the same result after being asked the same question.
https://www.longformexplainers.com/llm-inference/10-sampling
https://www.ibm.com/docs/en/watsonx/saas?topic=prompts-model...
thank god, these parameters are so confusing
> 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.
It is guaranteed to work worse than top_k=1, that’s for sure.
Depends on what you mean by determinism.
fwiw sonnet-5 also drops temperature (sonne-4 had it)
Good. These have been basically useless for the past few generations of models, and most of the time made the model perform worst.
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.