Hi there,
I'm facing an issue in using ner.openai.correct. I tried using the argument --model to select an openai model with high token limit (default is "text-davinci-003" which allows only 4000 tokens). It seems that the options are limited to GPT-3 completion models with low token limit.
Is it possible to use more powerful LLMs instead (e.g., gpt-3.5-turbo-16k) ?
Best,
Agapi
Hi @agapi.ris,
We're working on integrating newer models with our new spacy-llm. The code for this integration should be done in the next couple weeks and will be released as a part of Prodigy v1.13.
For now, this code is limited to the OpenAI Legacy Completions endpoint so the newer models are not available.
If you want to use the newer models now to help with Prodigy integration, you can follow the spacy-llm docs to generate LLM NER predictions for a set of data, save that data to Prodigy's JSONL format and correct those annotations using the built-in ner.manual recipe.
Otherwise, stay tuned for v1.13