No instruct dp #20
No instruct dp #20
Conversation
|
|
||
| # Build one pattern line per required timestamp: | ||
| # (YYYY-MM-DD HH:MM:SS,[-+]?\d+(?:\.\d+)?) | ||
| # No spaces allowed anywhere, so everything is literally "fixed" | ||
| # except for the numeric portion. | ||
| lines = [ | ||
| rf"\({re.escape(ts)},[-+]?\d{1,20}(?:\.\d{0,20})?\)" | ||
| for ts in required_timestamps | ||
| ] | ||
|
|
||
| # Join lines with exactly one "\n". | ||
| body = r"\n".join(lines) | ||
|
|
||
| # Return the full pattern, ensuring a single newline | ||
| # after <forecast> and before </forecast>. | ||
| return rf"<forecast>\n{body}\n</forecast>" |
There was a problem hiding this comment.
Is there a need to replace the original function? This might break reproducibility.
There was a problem hiding this comment.
The original function allowed for returns such as \n \n \n \n \n \n \n ... so I fixed it to this when working with Hymba. I think it would only break reproducibility if this code is somehow encoded into controlling the RNG for the llm or for the tasks, right?
There was a problem hiding this comment.
@marcotet what are your thoughts?
There was a problem hiding this comment.
Hmm, all results of our paper were with the previous code so I guess it was fine - we didn't get any errors.
If you think it's better maybe good to reproduce a model's results on CiK to confirm it doesn't break reproducibility. What do you think?
There was a problem hiding this comment.
Does mamba need this change?
| @@ -19,5 +19,6 @@ termcolor | |||
| tenacity | |||
| h5py | |||
| transformers>4.4.1 | |||
| tokenizers | |||
There was a problem hiding this comment.
Was this necessary for mamba or something?
There was a problem hiding this comment.
Yeah, I can remove it though since I didn't add the mamba models
There was a problem hiding this comment.
Yeah, that'd be great cause it might break the already-fragile requirements otherwise :D Unless you tested that it didn't :)
|
By the way, I'm testing if the code works for other base models. Will get back on that. |
Adds support for using Direct Prompt for models that are do not support instructions in HF.