param length_penalty effect on inference

[FatihcanUslu]

I fine-tuned a Coqui model, and I’m encountering an issue where very short input words cause the model to generate excessively long outputs containing repetitive “uhhhh” patterns and mostly unintelligible content. Model is good at generating medium or long input words.

To mitigate this, I attempted adjusting several config parameters:

      "length_penalty": -1.0,
      "repetition_penalty": 100.0,
      "top_k": 1,
      "top_p": 0.85,
      "num_gpt_outputs": 34

However, I am not seeing any noticeable effect from changing length_penalty or num_gpt_outputs. I tested some values (from -100 to +100), but the results remain essentially the same. So does length_penalty or num_gpt_outputs work?

The documentation (https://docs.coqui.ai/en/latest/models/xtts.html) provides two different definitions of length_penalty:

“A length penalty applied to the autoregressive decoder. Higher settings cause the model to produce more terse outputs. Defaults to 1.0.”

“An exponential penalty applied during beam-based generation. It is used as an exponent of the sequence length to divide the score. Since the score represents the (negative) log-likelihood, length_penalty > 0.0 promotes longer sequences, while length_penalty < 0.0 encourages shorter sequences.”

Also if i use high value it sends this warning:

The following generation flags are not valid and may be ignored: ['length_penalty']. Set TRANSFORMERS_VERBOSITY=info for more details.

[eginhard]

Thank you! I opened a new issue for this in #522 because that's a bug indeed.