Welcome to Software Development on Codidact!
Will you help us build our independent community of developers helping developers? We're small and trying to grow. We welcome questions about all aspects of software development, from design to code to QA and more. Got questions? Got answers? Got code you'd like someone to review? Please join us.
How to generate multi-line completions (code generation) with OpenAI?
I have finally received access to a trial account on OpenAI and I took it for a spin. My interest is with code generation, so first tried something simple using the playground:
Q: generate a standard model for Foo
A: public class Foo
{
public int Id { get; set; }
public DateTime DateCreation { get; set; }
public DateTime? DateModification { get; set; }
public int User { get; set; }
public bool Archive { get; set; }
}
Q: create a standard model for Bar
A: public class Bar
{
public int Id { get; set; }
public DateTime DateCreation { get; set; }
public DateTime? DateModification { get; set; }
public int User { get; set; }
public bool Archive { get; set; }
}
Q: create a standard model for MyNewModel
A: public class MyNewModel
{
public int Id { get; set; }
public DateTime DateCreation { get; set; }
public DateTime? DateModification { get; set; }
public int User { get; set; }
public bool Archive { get; set; }
}
(the last answer is automatically generated)
Playground parameters are as follows:
Engine = curie (second best after davinci, good results, faster and cheaper)
Temperature = 0 (always the same output for the same input)
Stop sequences = "A:" and "Q:"
Inject start text = "A:"
Inject restart text = "Q:"
This works as expected and seems to rely on create completion via GET.
I switched to code and found OpenAI-API-dotnet which uses the recommended create completion POST API method. My entire code is as follows:
var api = new OpenAIAPI(engine: Engine.Curie);
string text = @"
Q: generate a standard model for Foo
A: public class Foo
{
public int Id { get; set; }
public DateTime DateCreation { get; set; }
public DateTime? DateModification { get; set; }
public int User { get; set; }
public bool Archive { get; set; }
}
Q: create a standard model for Bar
A: public class Bar
{
public int Id { get; set; }
public DateTime DateCreation { get; set; }
public DateTime? DateModification { get; set; }
public int User { get; set; }
public bool Archive { get; set; }
}
Q: create a standard model for MyNewModel";
var textList = text.Split(new[] {"Q:", "A:"}, StringSplitOptions.RemoveEmptyEntries);
var prompt = string.Join("|endoftext|", textList);
var result = await api.Completions.CreateCompletionAsync(
prompt, 1024, 0, null, null, null, null,
null, null);
var splits = result.ToString().Split("|endoftext|", StringSplitOptions.RemoveEmptyEntries);
foreach (var line in splits)
{
Console.WriteLine(line);
}
The problem is that the engine seems to take more time to process (and costs more) almost the same input and also does not return any new lines. The output is the following:
|endoftext| public class MyNewModel
{ public int Id { get; set; } public DateTime DateCreation { get; set; } public DateTime? DateModification { get; set; } public int User { get; set; } public bool Archive { get; set; } }
|endoftext|
|endoftext|
|endoftext|
|endoftext|
........... (many more)
The actual generated code is almost what I need, but the new lines are missing. The engine needs to know which is the input and the output for each case, but the docs only mentions |endoftext| separator (I guess it should be used between the cases)
I have checked the API examples, but they only contain trivial examples (i.e. a few words on the same line).
Any idea how to correctly send the text to obtain a similar output as in the playground?
1 answer
I have reached OpenAI's support and one way to generate more accurate multiline responses is to use a clear separator between each question and answer block of text. My final working code is the following (### is the separator I have used):
string text =
@"Q: generate a standard model for Foo
A: public class Foo
{
public int Id { get; set; }
public DateTime DateCreation { get; set; }
public DateTime? DateModification { get; set; }
public int User { get; set; }
public bool Archive { get; set; }
}
###
Q: generate a standard model for Bar
A: public class Bar
{
public int Id { get; set; }
public DateTime DateCreation { get; set; }
public DateTime? DateModification { get; set; }
public int User { get; set; }
public bool Archive { get; set; }
}
###
Q: generate a standard model for MyNewModel";
// normalizing newlines
text = text.Replace(Environment.NewLine, "\n");
var result = await api.Completions.CreateCompletionAsync(
text, 1024, 0, 1, null, 0, 0,
null, null, "###");
Console.WriteLine(result);
The output is:
A: public class MyNewModel
{
public int Id { get; set; }
public DateTime DateCreation { get; set; }
public DateTime? DateModification { get; set; }
public int User { get; set; }
public bool Archive { get; set; }
}
For unknown reasons, I seem to get better formatting when working with the API (correct newlines) than when working with the playground.
Note: this answer was made possible by Adam Rhodes and it was also posted on OpenAI's community
0 comment threads
0 comment threads