Does generate's max_length influence training?

🤗Transformers
1

I’m finetuning a BART model from this tutorial: Google Colab and having a lot of trouble figuring out if I trained my model for the right document length.

I truncated my training data to 1028 tokens, but the default for generation is 12 tokens. If I bump this up to 512, will it still generate meaningful sentences? Is there another parameter I need to add to Seq2SeqTrainingArguments? Do the GenerationConfig parameters have any impact on training?

The BartConfig has the following section on task_specific_params.

  "task_specific_params": {
    "summarization": {
      "length_penalty": 1.0,
      "max_length": 128,
      "min_length": 12,
      "num_beams": 4
    },
    "summarization_cnn": {
      "length_penalty": 2.0,
      "max_length": 142,
      "min_length": 56,
      "num_beams": 4
    },
    "summarization_xsum": {
      "length_penalty": 1.0,
      "max_length": 62,
      "min_length": 11,
      "num_beams": 6
    }

Thanks!