research idea pad

here are some ideas i or someone else can check out

• preliminary experiments by a friend of mine shows that vicuna-13b-free is able to better follow the initial prompt better than vicuna-13b 1.0, is this purely because of the ethical restraint? are these models the SAME exact dataset? how were they trained, is free edition a finetune on 1.0, or something else?
  • resources required: a gpu capable of running vicuna-13b-free (3060 12gb fits)
• decentralized model distribution projects
  • torrent is a good starter to build these on
  • never really warmed up to IPFS
  • ggml file is unstable, don't submit it to torrents imo
    • gguf is a somewhat stable now! %at=2024-06-16T22:32:57.662Z
    • https://github.com/ggerganov/ggml/pull/302
  • huggingface2torrent
  • civitai2torrent
  • resources required:
    • code
    • a fuckton of storage
  • unanswered question: how to index these torrents?
• how to train models in a distributed manner across random gpus through the world? %at=2023-06-22T23:22:37