generation_config.json

{
  "bos_token_id": 1,
  "eos_token_id": 2,
  "pad_token_id": 2,
  "max_new_tokens": 512,
  "max_length": 4096,
  "do_sample": true,
  "temperature": 0.7,
  "top_p": 0.9,
  "top_k": 50,
  "repetition_penalty": 1.1,
  "no_repeat_ngram_size": 3,
  "_generation_notes": "Use max_new_tokens instead of max_length for best results"
}