`general.name` in GGUF metadata is set to a random-looking sequence of characters

#2
by rotemdan - opened

In the GGUF metadata for LFM2-24B-A2B the general.name property is set to:

Ef218A605E23739A1302869B9B3618C8B3F7Eb0D

This property is used as the model's display name on some backends.

Other models I looked at

I also looked on some other GGUF versions of LFM models on this repository:

In model LFM2.5-Audio-1.5Bit's set to a human-readable value:

LFM2.5 Audio 1.5B

But in model LFM2.5-1.2B-Thinking it's also set to random characters:

E7Ae5Ebc4Be00163B1D1Bd366569776D85957F53

In LFM2-2.6B-Transcript it is:

95d6f2b251c506c467d00d4f50ea41067f363a3f

In LFM2.5-1.2B-JP:

B2B2Fcaf821De488Ac59Fb36353488341Bb61E0C

Anyway, I guess didn't report it sooner because I previously only used the GGUF files from Unsloth, which have human-readable names in general.name.

Liquid AI org

Thanks for reporting @rotemdan , the conversion script derives it heuristically and can fall back to the directory name in the HF cache. Will update the conversion procedure to explicitly set general.name via --model-name flag.

Sign up or log in to comment