generation_config.json

{
  "_from_model_config": true,
  "bos_token_id": 1,
  "eos_token_id": 2,
  "pad_token_id": 0,
  "transformers_version": "4.35.2",
  
  "max_length": 512,            // Limits the length of the output text
  "temperature": 0.7,           // Controls randomness, higher means more random
  "top_p": 0.9,                 // Nucleus sampling, considers only top_p highest probability tokens
  "top_k": 50,                  // Limits the sample space to top_k tokens with highest probabilities
  "do_sample": true,            // Enables sampling instead of greedy decoding
  "repetition_penalty": 1.2,    // Penalizes repetition
  "num_return_sequences": 1     // Number of output sequences to generate
}