Inquiries about Prodigy 1.12 and Future GPT-4 Integration

, ,
1

Hello Prodigy Support Team,

I hope this message finds you well. I have recently upgraded to Prodigy 1.12 and have been enjoying the new features and enhancements. I had a couple of questions that I was hoping you could assist me with.

  1. Can we use GPT-4 in Prodigy recipes? If not, do you have any information on when GPT-4 will be available for integration in Prodigy recipes? I'm eager to explore the capabilities of the newer GPT version and would appreciate any insights or updates on this matter.
  2. I have been utilizing the "rel.manual" recipe in Prodigy to build relations, and it has been a great experience. However, I was wondering if there is currently a way to train the relation model using the annotated data and obtain predictions on unseen data. If not, I kindly request that you consider implementing a feature that allows us to easily train our relation models and generate predictions on new data. This would greatly enhance the practicality and usefulness of Prodigy in relation-based tasks.

Thank you for your time and support. I look forward to hearing from you and appreciate any information or guidance you can provide regarding my inquiries.

Best regards,

2

Hi there!

  1. The Prodigy recipes should integrate with the non-chat based models from OpenAI. My understanding is that OpenAI recently declared these endpoints to be "legacy" and will favor the chat completion based models via the https://api.openai.com/v1/chat/completions endpoint. The plan is to support these chat based models for sure, but the integration will occur via spacy-llm. This way, you can treat the LLM powered predictions just like a "normal" spaCy pipeline and Prodigy will just integrate with spacy-llm directly to support a wide range of LLM models. We're working on getting support for spacy-llm done and I expect it won't take long for us to release v1.13.

  2. Funny that you mention this because spacy-llm actually has support for this! Support for relationship models via spacy-llm models is on the roadmap for v1.13 for sure. You might also be able to prodigy train to train a coref model (as of v1.12) using spacy-experimental. This could be re-used in a custom recipe too.

1 Like