I performed a silly experiment by posing the question "Which animal do you identify with?" to different LLMs. Curiously there are very specific preferences depending on the LLM family, which are likely the result of different finetuning datasets.

None of them identified as llama...

I expect that a number of similar questions could be used to build an LLM fingerprinting evaluation dataset. I wonder whether it is possible to make it resilient to finetuning.

Heatmeap of responses at default temperature (n=10)

Heatmeap of responses at zero temperature (n=3)

You can find the evaluation scripts here:
https://github.com/cpldcpu/llmfingerprint