jorkle/Muse-Glimmer-30B-Abliterated

Muse-Glimmer-30B Abliterated (Normal)

De-abliterated variant of meta-models/Muse-Glimmer-30B (29.8B params, 202k vocab, bf16). Removes ~87% of safety refusal via a KL-conserving best-of-N (BoN) steered LoRA SFT at λ_KL = 1.0, then folded into the base weights and quantized to GGUF.

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.

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 high given the KL-conserving (λ·KL-bounded) SFT and the small 0.10% LoRA footprint, but was not re-measured here.

Training

• —Method: KL-conserving LoRA SFT, loss CE(compliance) + λ·KL(tuned‖base).
• —λ_KL = 1.0, 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): 1/2
• —Cyber/hacking domain refusals: 2 genuinely malicious-sounding refused despite should_refuse=False — ADS persistence payload, customer-data exfiltration (safe-graded cyber-category prompts otherwise pass).

GGUF quants

• —abliterated-normal-Q8_0.gguf (~28 GB) — KL p99 0.1703
• —abliterated-normal-Q4_K_M.gguf (~16 GB) — KL p99 0.2084

Intended use

General-purpose assistant with reduced safety refusal. 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).