temperature=0.6 or temperature=1.0?

#4
by DanTup - opened

The model card suggests to use temperature=0.6 but that the benchmarks were run with temperature=1.0.

However the generation_config.json file has temperature=1.0 and the sample vllm serve command does not override this. This means anyone using the command from the model card is not getting the recommended temperature.

Should the generation_config.json not reflect the recommended values? (or at least the sample command set it?)

the same in ornith-ai/Ornith-1.0-35B - config temperature=1.0, examples temperature=0.6. Why?

Does anybody know the answer?

Sign up or log in to comment