Inconsistency in responses using gemma3 12B model

[shivanis1406]

In the second stage, the generative model converts these distributions into actual text responses through one of several decoding strategies. A simple decoding strategy might select the most likely token at every timestep. This process would always be deterministic. However, you could instead choose to generate a response by randomly sampling over the distribution returned by the model. This process would be stochastic (random). Control the degree of randomness allowed in this decoding process by setting the temperature. A temperature of 0 means only the most likely tokens are selected, and there's no randomness. Conversely, a high temperature injects a high degree of randomness into the tokens selected by the model, leading to more unexpected, surprising model responses.
https://ai.google.dev/gemini-api/docs/prompting-strategies

Does above hold true for gemma as well? Why does unsloth suggest temperature = 1 ?

https://docs.unsloth.ai/basics/gemma-3n-how-to-run-and-fine-tune