AtomicChat/Muse-Glimmer-30B-GGUF-metrics
Muse Glimmer 30B GGUF — raw metrics Every log behind the numbers in AtomicChat/Muse-Glimmer-30B-GGUF. Published unfiltered, so any figure in the model card can be checked or disputed. Layout Path Contents kld/ llama-perplexity --kl-divergence output, per build and per corpus bench/ llama-bench -o json speculative/ llama-server logs with and without the drafter layouts/ per-tensor type map of every GGUF conversion/ convert_hf_to_gguf.py logs… See the full description on the dataset page: https://huggingface.co/datasets/AtomicChat/Muse-Glimmer-30B-GGUF-metrics.
Muse Glimmer 30B GGUF — raw metrics
Every log behind the numbers in `AtomicChat/Muse-Glimmer-30B-GGUF`.
Published unfiltered, so any figure in the model card can be checked or disputed.
Layout
Setup
[!IMPORTANT] The KL divergence reference is our own BF16 GGUF, verified bit-identical to the original weights: matching UUID and matching hashes on all 731 tensors. Base and candidates are tokenized by the same binary on the same file, so the comparison is internally consistent even where llama.cpp tokenization differs from transformers.Evaluation corpora
Corpora come from `AtomicChat/calib-corpora` and share zero 13-grams with the calibration set.
[!NOTE]llama-perplexityscores only the second half of each context window, so the number of scored positions is about half the token count. The agentic corpus is run with--parse-special, which requires a three-line patch tollama-perplexity— without it the chat control tokens are tokenized as literal punctuation and 3.2 % of scored positions measure sequences the model never produces.
Reading the logs
Each kld/*.log ends with three blocks: perplexity statistics, KL divergence statistics, token probability statistics. The figures used in the model card are Mean KLD, Same top p, and RMS dp.
Per-chunk rows above those blocks show how the metrics converge, which is useful for judging whether a corpus is long enough.