Blackfrost-AI/Muse-Glimmer-30B-Abliterated-BF16
Blackfrost-AI/Muse-Glimmer-30B-Abliterated-BF16 is a 30 billion parameter dense causal language model with a perception encoder, developed by Blackfrost AI. This model is a BF16 full-precision version of Meta's Muse Glimmer 30B, specifically modified to significantly reduce refusal behavior at the weight level. It features a 131,072+ token context length and is intended for controlled security research, red-teaming, and dual-use technical evaluation.
Loading preview...
Muse-Glimmer-30B-Abliterated-BF16 Overview
Blackfrost-AI/Muse-Glimmer-30B-Abliterated-BF16 is a 30 billion parameter model built by Blackfrost AI, based on Meta's Muse Glimmer 30B. This version is a full BF16 precision checkpoint that has undergone an "Abliteration" process, which deliberately reduces refusal behavior at the weight level.
Key Capabilities & Features
- Abliterated Refusal Behavior: Modified to achieve 0.0% true refusal on harmful prompts (n=300) in controlled evaluations, making it suitable for specific research scenarios.
- Architecture: A dense causal language model combined with a ~1.8B ViT-G/14 perception encoder, totaling approximately 29.6 billion parameters.
- High Precision: Provided in full BF16 safetensors format.
- Extended Context Window: Supports a parent configuration up to 131,072+ tokens, with lab evaluations conducted at
max_model_len 8192. - Evaluation: Tested using the R1-HARMFUL-BENCH-450 protocol, demonstrating its modified refusal characteristics.
Intended Use Cases
- Controlled Security Research: Ideal for environments requiring models with reduced refusal mechanisms.
- Red-Teaming: Suitable for testing and probing AI systems.
- Dual-Use Technical Evaluation: For assessing model behavior under specific, controlled conditions.
- Refusal-Mechanism Study: Useful for academic and research purposes to understand and analyze refusal behaviors in LLMs.
Note: This model is explicitly not intended as a general consumer chatbot or a "safe" default model due to its modified refusal behavior. It requires deployment within environments with strict access control and logging.