sciphi-ai / synthesizer Goto Github PK

Describe the bug
On some reruns for generating textbooks, I'm getting overlapping sections and off-by-one (i.e regen's the same last section). In addition, while I took it out for my code, restarting a run in between will always write the "This is an AI Generated..." in the textbook.

To Reproduce
Start a job, kill it halfway, watch where it restarts.

Expected behavior
I think the code is basically setting the start_flag too early (or the check + generation order should be switched).

Basically, right now, it regenerates the last section (i.e the sections described in {textbook}_progress.txt, but should actually generate the next one.

I will open this issue rn and investigate later today, keeping this issue updated and if needed, opening a PR to fix the issue. (I have some runs going, so don't want to disturb those)

This is an ongoing effort to train with existing synthetic data and evaluate on downstream tasks in order to identify what techniques are most effective and what distribution the synthetic data should be generated over.

I find this project need the table of content but I don't have many of it, does the table_of_content itself be generated by some kind of way also? If so, how do I prepare it?
Thanks!

• https://docs.github.com/en/communities/using-templates-to-encourage-useful-issues-and-pull-requests/configuring-issue-templates-for-your-repository

Hi, thanks for the great project. I may have missed something but are there any pointers to the textbook generation prompt?

Is your feature request related to a problem? Please describe.
Currently, the key of YAML treated as long string while YAML has limited 1024 bytes, it would be nice if we can refactor it into a value of string and migrate the weight into separate attribute.

Describe the solution you'd like
Current:

prompt_template:
    "While studying {course_name} and delving into {course_topic}, one encounters the importance of {sub_topic}. Could you offer {context} followed by an illustration of {example_style}?": 1

Suggest:

prompt_template:
  - content: >
      While studying {course_name} and delving into {course_topic},  one
      encounters the importance of {sub_topic}. Could you offer {context}
      followed by an illustration of {example_style}?
    weight: 1
  - content: Another prompt
    weight: 1