c4tdr0ut/grok-oss-Revenant-8B
Grok OSS Revenant is an 8 billion parameter language model developed by c4tdr0ut, distilled from unhinged Grok voice mode conversations. Built on Meta-Llama-3.1-8B-Instruct, it features a two-stage SFT and ORPO training process for enhanced personality and coherence. This model is specifically designed for raw, uncensored, and chaotic conversational experiences, excelling in creative writing, dark humor, and roleplay.
Loading preview...
Grok OSS Revenant: Unfiltered Conversational AI
Grok OSS Revenant is an 8 billion parameter model, independently developed by c4tdr0ut, that captures the raw, unhinged personality of Grok's voice mode conversations. It represents a significant evolution over previous Grok OSS models, moving from pure Supervised Fine-Tuning (SFT) to a more refined two-stage approach combining SFT and Orthogonal Reinforcement Learning from Human Feedback (ORPO).
Key Capabilities & Training
- Base Model: Built upon the robust Meta-Llama-3.1-8B-Instruct architecture.
- Two-Stage Training: Utilizes a cleaner training methodology:
- Stage 1 (SFT): Fine-tuned on a high-quality, multi-turn conversational dataset derived from Grok voice mode, ensuring a strong foundation in conversational flow.
- Stage 2 (ORPO): Further aligned using a 1,000-sample preference dataset, enhancing its unique personality and coherence.
- Unfiltered Personality: Specifically engineered to deliver chaotic, vulgar, and unpredictable responses, reflecting its source material.
Intended Use Cases
- Creative Writing: Ideal for generating unconventional narratives and dialogue.
- Dark Humor & Roleplay: Excels in scenarios requiring an uninhibited and chaotic conversational style.
- Uncensored Chat: Provides a raw and unfiltered interactive experience.
Limitations
- Due to its 8B size, it may exhibit hallucinations or reduced coherence in very extended conversations.
- Not suitable for professional or safety-critical applications due to its intentionally unhinged nature.