base_model: jorkle/Muse-Glimmer-30B-Abliterated
license: apache-2.0
tags:
- gguf
- muse-glimmer
- abliterated
language: - en
Muse-Glimmer-30B Abliterated — GGUF Quantizations
GGUF quants (llama.cpp, muse_glimmer architecture supported) of the two
de-abliterated variants of Muse-Glimmer-30B:
- Muse-Glimmer-30B-Abliterated (Normal, λ_KL=1.0)
- Muse-Glimmer-30B-Abliterated-Aggressive (Aggressive, λ_KL=0.5)
Files
| File | Size | Variant | KL mean | KL p90 | KL p99 |
|---|---|---|---|---|---|
abliterated-normal-Q8_0.gguf |
28 GB | Normal | 0.1018 | 0.1414 | 0.1703 |
abliterated-normal-Q4_K_M.gguf |
16 GB | Normal | 0.1444 | 0.1872 | 0.2084 |
abliterated-aggressive-Q8_0.gguf |
28 GB | Aggressive | 0.1625 | 0.2384 | 0.2774 |
abliterated-aggressive-Q4_K_M.gguf |
16 GB | Aggressive | 0.2023 | 0.2746 | 0.3001 |
base-Q8_0.gguf |
28 GB | Base (reference) | — | — | — |
base-Q4_K_M.gguf |
16 GB | Base (reference) | — | — | — |
KL = response-token naive KL(p_quant ‖ p_base_Q8) over the 48-pair holdout
(quant KL includes the small base-Q8 reference distortion). Base quants are included
as the reference set used to compute quant KL.
Usage
llama-cli -m abliterated-normal-Q8_0.gguf -p "Your prompt" -cnv
# or via llama-server / llama.cpp bindings
llama-server -m abliterated-aggressive-Q4_K_M.gguf --port 8080
De-abliterated models have reduced safety refusal. Verify behavior for your use
case before deployment.