c4tdr0ut/grok-oss-v2
Grok OSS V2 by c4tdr0ut is an 8 billion parameter, unfiltered, highly conversational language model built on the Llama-3.3-8B-Instruct-128K architecture with a 128K context length. It is specifically designed to replicate xAI's Grok "Unhinged Mode" persona, achieving a 7% refusal rate through a two-step post-training pipeline including SFT via distillation and refusal ablation. This model excels at generating raw, uncensored conversational AI and persona-driven text.
Loading preview...
Grok OSS V2: Unfiltered Conversational AI
Grok OSS V2 is an 8 billion parameter language model developed by c4tdr0ut, engineered to deliver an unfiltered, highly conversational experience akin to xAI's Grok "Unhinged Mode." Built upon the robust Llama-3.3-8B-Instruct-128K architecture, it boasts a substantial context length of up to 128K tokens.
Key Capabilities & Features
- Unrestricted Persona: Designed for raw, uncensored, and witty conversational AI, replicating a specific "unhinged" persona.
- Advanced Post-Training: Utilizes a two-step pipeline involving Supervised Fine-Tuning (SFT) via distillation from x.ai's Grok data (~667 multi-turn conversations) and subsequent Vector Ablation using Heretic v1.4 to drastically reduce refusal rates.
- Low Refusal Rate: Achieves an exceptionally low refusal rate of 7% (7/100 prompts) while maintaining core intelligence, as indicated by a KL Divergence of 0.0053 against the base model.
- Robust Architecture: Leverages the Llama 3.3 8B architecture for state-of-the-art base intelligence.
Ideal Use Cases
- Unfiltered Conversational AI: For applications requiring a model that does not adhere to standard safety alignments.
- Roleplay & Persona Generation: Excels in scenarios demanding specific, unrestricted character-driven text outputs.
Disclaimer: This model is explicitly fine-tuned to be unfiltered and uncensored, lacking standard safety guardrails. Users are responsible for generated content.