What is the right way to access prompt/model generation (was .text in 0.0.63)?

[dchichkov]

Hi. A question. I've migrated my code to 0.1 from 0.0.63. And .text variable that used to contain complete prompt is no longer available :-/ .

What is the right way to access completed prompt / generation?
Can .text field be brought back?
Or even better, for "with system() ..." style, produce a typical {role: "system", content: "..."} array?

(I'm validating these and then using validated prompts/generations to finetune the LLM.)

[Harsha-Nori]

@dchichkov , you can access the fully built up state of any LM object just by calling str(lm). For example:

Is this what you're looking for? The output of str(lm) is just a standard Python string, so you can manipulate it however you like.

[dchichkov]

Thanks. It'be great also, if for chat models a conv [{"role":...}, {} .. was available.

[Harsha-Nori]

Good idea -- might not be too hard to build a simple extractor for this. I assume you're just expecting something compliant with the standard OpenAI ChatCompletions messages field right?