Text Generation
Transformers
PyTorch
English
gpt_neox
text-generation-inference
Inference Endpoints

Unwanted repetitive response

#12
by sdranju - opened

I always get repetitive response and it's just endless.

Setting pad_token_id to eos_token_id:0 for open-end generation.
Paris

Q: What is the capital of the U.S. state of California?
A: Sacramento

Q: What is the name of the country that has the largest population in Europe?
A: Russia

Q: What is the capital of the country that has the largest population in Europe?
A: Moscow

Q: What is the capital of the country that has the largest population in Europe?
A: Moscow

Q: What is the capital of the country that has the largest population in Europe?
A: Moscow

Together org

Hi @sdranju , thanks for your feedback! Can you let me know what prompt template and generation parameters you are using?

Hi, same for me if I ask the model the model to generate questions based on a paragraph:
(I am not very successful at making it follow any instructions)

My prompt:
{paragraph}
Generate questions for a quiz:

  1. What is Machine Learning?
  2. Why...

Response:
{paragraph}
Generate questions for a quiz:

  1. What is Machine Learning?
  2. Why are neural networks used?
  3. What are adversarial networks?
  4. What are adversarial networks?
  5. What are adversarial networks?
  6. What are adversarial networks?
    ...
Together org

Hi @Michelangiolo , did you try to set the parameter repetition_penalty to something > 1.0? Setting it to e.g. 1.2 can help in eliminating repetitions (see this paper for details).

Sign up or log in to comment