Sampling parameters recommendation
Quick question: I'm still as much a fan of your models, but taking a closer look, I realize you rarely provide recommendations on generation sampling parameters.
I think the truth lies somewhere between the two parent models, Qwen and Gemma, who are the origin of the branch:
gemma-4-E4B:
- temperature=1.0
- top_p=0.95
- top_k=64
Qwen3.5-4B (Thinking mode for general tasks)
- temperature=1.0
- top_p=0.95
- top_k=20
- min_p=0.0
- presence_penalty=1.5
- repetition_penalty=1.0
Could you share any recommendations on this topic?
Thanks for the kind words and the good question, @Cheuv !
Genesis runs on the Gemma 4 architecture (the father's attention stack is kept intact), so we recommend starting from the Gemma 4 defaults, which are also what ships in generation_config.json:
temperature=1.0
top_p=0.95
top_k=64
For benchmarks and tasks with a single correct answer, greedy decoding (do_sample=False) is what we used in our own evaluations. If you see repetition in very long outputs, a mild presence_penalty (around 0.5 to 1.0) is a reasonable thing to try, but we have not run a formal sweep on it, so please treat that as a starting point rather than a tuned value.
Thanks again for using the model!