GGUF-NEW-METADATA(1)

πŸ“¦
version 10438 Β· 2 versions
, in ubuntu stonking

✨ NAME ✨

gguf-new-metadata β€” bestie, it's literally gguf-new-metadata. That's the name. That's the whole vibe. πŸ’…


πŸ¦‹ DESCRIPTION πŸ¦‹

okay so the actual usage goes a lil something like this:

usage: gguf-new-metadata [-h] [--general-name "name"]
      [--general-description "Description ..."]
      [--chat-template "{% ... %} ..."]  [--chat-template-config
      tokenizer_config.json] [--chat-template-file chat_template.jinja]
      [--pre-tokenizer "pre tokenizer"] [--remove-metadata general.url]
      [--special-token bos | eos | eot | eom | unknown | seperator |
      padding | mask | fim_pre | fim_suf | fim_mid | fim_pad | fim_rep |
      fim_sep | prefix | suffix | middle "<token>"]
      [--special-token-by-id bos | eos | eot | eom | unknown |
      seperator | padding | mask | fim_pre | fim_suf | fim_mid |
      fim_pad | fim_rep | fim_sep | prefix | suffix | middle 0]
      [--force] [--verbose] input output

And what does it do, you ask? Make a copy of a GGUF file with new metadata. That's it. That's the whole job. Glow-up for your metadata, no cap. 🫢


πŸŽ€ positional arguments:

  • input β€” the GGUF format model input filename, aka the file you're copying from, the before picture πŸ“Έ
  • output β€” the GGUF format model output filename, aka where the glow-up lands, the after picture ✨

πŸ’– options:

  • -h, --help β€” show this help message and exit. the classic "wait what does this do again" button πŸ™ˆ
  • --general-name "name" β€” sets the model's general.name πŸ’«
  • --general-description "Description ..." β€” sets the model's general.description, so go off and describe your model, bestie ✍️
  • --chat-template "{% ... %} ..." β€” a chat template string (or a JSON string containing templates) πŸ—£οΈ
  • --chat-template-config tokenizer_config.json β€” a config file containing chat template(s) πŸ“„
  • --chat-template-file chat_template.jinja β€” a Jinja file containing a chat template 🧚
  • --pre-tokenizer "pre tokenizer" β€” sets the model's tokenizer.ggml.pre πŸ”€
  • --remove-metadata general.url β€” removes metadata from the output model, and you do it by key name πŸ—‘οΈ
  • --special-token bos | eos | eot | eom | unknown | seperator | padding | mask | fim_pre | fim_suf | fim_mid | fim_pad | fim_rep | fim_sep | prefix | suffix | middle "<token>" β€” special token, specified by value πŸ‘‘
  • --special-token-by-id bos | eos | eot | eom | unknown | seperator | padding | mask | fim_pre | fim_suf | fim_mid | fim_pad | fim_rep | fim_sep | prefix | suffix | middle 0 β€” special token, but specified by id this time πŸ”’
  • --force β€” bypass warnings without confirmation. careful with this one, bestie, it's the "yes I'm sure, stop asking me" button and it means exactly that πŸ’€
  • --verbose β€” increase output verbosity, for when you want the program to tell you everything it's feeling πŸ—―οΈ
ngl the token-type list has seperator spelled like that in the original, and we are not fixing it, because copying things character for character is the whole assignment πŸ’…

that's the whole page, babe. copy a GGUF, swap the metadata, hit --force only when you mean it, and go be the main character of your own model repo. 🌈 ⭐

β€” brought to you by debian, dated September 2026, GGUF-NEW-METADATA(1), User Commands section, mwah 😘 πŸ’–

πŸ“¦ where this page comes from