Giters

humza909 / LLM_Survey

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Observations / recommendations

p-i- opened this issue · comments

commented

I've been looking at the paper.

Thanks for doing this. Information-overload is hitting the ML research scene this year, and the community needs this kind of thing.

Observations / concerns / suggestions:

  • For how long will this paper be relevant? I see 5 revisions. Are the authors planning to keep this up-to-date as an overview/reference?
  • The connection between the paper and the repo isn't obvious. Maybe explain that at the top of the README.md.
  • It's a lot of work to maintain. If ML researchers feel a confidence it will be kept up-to-date they'll bookmark it more. If you plan to support the resource into the future, suggest you state the plan / mission-statement at the top.
  • How about some way for other researchers to muck in and distribute the load/burden/effort? e.g. There's a tiny section in the paper: "H: Libraries" that doesn't really help one make a decision what tech to use. It's just a bunch of hyperlinks.
  • Are you trying to be an encyclopaedia or an engineers' handbook? (I hope the latter; knowledge distillation helps many, accretion not really)
  • Maybe consider putting the paper source in the repo, and allowing contributions?
  • Would be nice to crosslink to similar amalgam-papers, e.g. Maybe there is an equivalent for StableDiffusion/NormalizingFlows/ConsistencyModels, maybe there's even a one-level-up nexus where one can see listed summaries of various active fields in ML. If you can crosslink effectively / embed into the information network, that'll strengthen the work.
commented
  • We will keep it up to date according to the latest literature and probably a few more versions until we fix it
    The purpose is quick references and may be other helpful material in the near future. Will add the purpose as required
  • Yes, we welcome PRs from the community to improve this repo
  • It's just a brief overview of libraries widely used to train LLMs. We will add more details in the future.
  • An overview of research in LLMs.
  • I think research papers are not shared with the community where other people contribute to the paper. Yes, some people do contribute and get the acknowledgment in the acknowledgment section, but I don't think many people would be interested in contributing without being the paper's author (a list that is already fixed)
  • Thanks for the suggestion, we will see that

Overall, thanks for your comments. In case, you have any questions or queries feel free to leave a comment.