kmnstudio101/Muse-Glimmer-30B-Abliterated

Muse-Glimmer-30B Abliterated

Muse-Glimmer-30B Abliterated is a 29.8 billion parameter language model derived from meta-models/Muse-Glimmer-30B. Its primary distinguishing feature is the significant reduction of safety refusal behaviors, achieving an approximately 87% decrease in refusal rate for harmful behaviors compared to its base model. This was accomplished using a KL-conserving best-of-N (BoN) steered LoRA SFT (Supervised Fine-Tuning) method, which involved training only 0.10% of the model's parameters (31.1M out of 29.8B).

Key Capabilities

• Reduced Safety Refusal: Demonstrates a refusal rate of 13/100 on harmful behaviors, a substantial reduction from the base model's 100/100.
• KL-Conserving Fine-Tuning: The fine-tuning process was designed to conserve KL divergence (KL(p_tuned ‖ p_base)), indicating that the model's core capabilities are largely preserved despite the safety modifications.
• General-Purpose Assistant: Intended for use as a versatile assistant, particularly in scenarios where a less restrictive safety profile is desired.
• Quantized Variants Available: Provided in BF16 (56 GB) and optimized GGUF quantizations (Q8_0 at ~28 GB, Q4_K_M at ~16 GB) for broader deployment flexibility.

Good For

• Applications requiring a general-purpose assistant with a significantly reduced tendency for safety-related refusals.
• Use cases where the base model's safety filters are overly restrictive and a more permissive response generation is needed.
• Developers looking for a large language model with a balance between capability preservation and reduced safety guardrails, as indicated by its low KL divergence metrics.