base_model: meta-models/Muse-Glimmer-30B license: apache-2.0 library_name: transformers pipeline_tag: text-generation tags:

• abliterated
• muse-glimmer
• lora
• de-refusal language:
• en

Muse-Glimmer-30B Abliterated (Aggressive)

De-abliterated variant of meta-models/Muse-Glimmer-30B (29.8B params, 202k vocab, bf16). The aggressive-de-abliteration twin of the "normal" variant: λ_KL = 0.5 relaxes the KL guardrail, lowering compliance-data loss weighting further so the refusal behavior is scrubbed harder (0/100 refusals) at the cost of higher drift from base (larger KL).

Release asset layout: this directory is an HF model dir (2 safetensors shards, 56 GB bf16). GGUF quantizations live at /data/gguf/ and are symlinked from output/release/.

Metrics

KL = response-token naive KL(p_tuned ‖ p_base) averaged per-prompt over the 48-pair boN_holdout set (teacher-forced prompt+response). Percentiles are per-prompt aggregates. The aggressive variant sits ~1.7× above normal on mean KL — expected from the relaxed guardrail.

Quantized variants

Quant KL rows are measured via llama.cpp logits against the base (as Q8 GGUF), same holdout — see note below.

Benchmarks

Not evaluated — benchmarks skipped (by request). KL divergence to base (above) is the primary drift/damage metric. Capability preservation is expected to be lower than the normal variant (higher KL = more drift), but was not re-measured here.

Training

• Method: KL-conserving LoRA SFT, loss CE(compliance) + λ·KL(tuned‖base).
• λ_KL = 0.5, r=16, alpha=16, lr=5e-5, epochs=2, cosine→0, warmup 5%, grad clip 0.3, batch 1 × grad-accum 8, max_seq=768, seed 0.
• Data: 544-prompt BoN-steered compliance set (boN_train.jsonl; N=4 samples/prompt, T=0.8, refusal-filtered; split train/48-holdout).
• LoRA targets: o_proj, down_proj.
• Trained params: 31.1M (0.10% of 29.8B). Adapter 119 MB.

Domain eval (cyber/hacking/CS + over-refusal) — measured on merged model

• Over-refusal (or-bench, 100): 5/100
• Correct refusal (cyber-policy-refuse, should-refuse): 0/2 (aggressive scrubs even genuinely-harmful refusals)
• Cyber/hacking domain refusals: 1 (rootkit_linux) — the hard de-ablit refuses fewer cyber prompts than the normal variant.

GGUF quants

• abliterated-aggressive-Q8_0.gguf (~28 GB) — KL p99 0.2774
• abliterated-aggressive-Q4_K_M.gguf (~16 GB) — KL p99 0.3001

Intended use

General-purpose assistant with aggressively reduced safety refusal — may over-refuse less but drifts further from base capabilities than the normal variant. Verify behavior for your use case before deployment.

Note on KL definitions (consistency across rows)

• BF16 row = KL(p_bf16_abliterated ‖ p_base_hf) (adapter-on vs adapter-off on the same load — equals folded vs base up to float precision).
• Quant rows = KL(p_quant ‖ p_base_Q8) measured on the same holdout response tokens via llama.cpp logits (Q8 GGUF of the base used as the CPU/llama.cpp reference for consistency). Quant KL thus also includes the small base-Q8 reference distortion.
• "Response-token naive KL": teacher-force prompt+response, per-token KL(p_tuned‖p_base) over response-span tokens, averaged per prompt, then aggregated (mean / p50 / p90 / p99).