Can We Enhance Output by Including an Example in the JSON Input Template?
Hi everyone,
I’ve been exploring ways to improve the quality of outputs in the models and had an idea:
Is it possible to include an example of the desired output directly in the JSON input template? or, could we embed an example in the prompt structure or another part of the input, so the model has a clearer understanding of the expected format?
Here’s an example of the current prompt structure:
prompts = [
f"""<|input|>\n### Template:\n{template}\n### Text:\n{text}\n\n<|output|>"""
for text in texts]
If we modify it to include a specific example of the desired output, would that help improve the quality of results?
Additionally, are there any other strategies or best practices to enhance this approach and generate better outputs?
Hello,
This was actually an exact feature of the version 1.0 models. You can try it out on your problem there to see if it helps, but ultimately we found it of limited use in most cases so removed it in 1.5.
However, we will also be releasing version 2.0 very soon which will be able to properly make use of full in-context examples, so keep an eye out for that release. :)