Reproducing results with local models on circle packing

[rahulmoorthy19]

Thank you for the excellent repository and the detailed setup instructions.

I’m trying to set up ShinkaEvolve with local models using Ollama. I ran it with qwen3:8b as the main LLM (and meta LLM), and qwen3-embedding:8b as the embedding model. However, even after ~150 generations with the default large_budget configuration (only changing the LLM and embedding models as mentioned), the best score I’ve been able to achieve is about 1.6.

I wanted to check if this might be a limitation of using an 8B parameter model for this task, or do you suspect something else?

Thank you!

[rahulmoorthy19]

The changes for local model are very similar to the pull request-

"added local model support, and minor fixes to work on ubuntu #33"

[zia1138]

@rahulmoorthy19 check out the change in #58
It's pretty important.

[rahulmoorthy19]

Thank you!