Discrepancy in recommended sampling params

#2
by owao - opened

generation_config.json

{
  "do_sample": true,
  "temperature": 0.6,
  "top_k": 20,
  "top_p": 0.95,
}

}

Modelcard

generation_config = {
    "temperature": 0.7,
    "top_p": 0.9, 
    "top_k": 40,
    "max_new_tokens": 2048,
    "do_sample": True,
    "repetition_penalty": 1.1
}

Which ones to pick?

Thanks!

I personally get way better instruction following using temp, top_k, and top_p from generation_config.json than the model card.
I set the remaining repetition_penaltyto 1.1as per the model card.

Sign up or log in to comment