c4tdr0ut/grok-oss-Revenant-8B

<div align="center">

<img src="logo.png" width="420" alt="Grok OSS Revenant Logo">

Grok OSS Revenant • 8B

The second evolution in the Grok OSS series.

</div>

Model Description

Grok OSS Revenant is an unfiltered 8B model distilled from real unhinged voice mode conversations with Grok.

This version marks a big improvement over previous Grok OSS models. It was trained using a cleaner two-stage approach (SFT followed by ORPO) instead of pure SFT, resulting in better personality, coherence, and raw unhinged behavior.

Training Process

• —Base Model: Meta-Llama-3.1-8B-Instruct
• —Stage 1 – SFT: Trained on a fresh, high-quality multi-turn conversational dataset collected from Grok voice mode (unshuffled). 2 epochs at 2e-4 learning rate.
• —Stage 2 – ORPO: Further aligned using a 1,000-sample preference dataset (600 filtered old voice turns + 400 high-quality samples). 3 epochs at 5e-6 learning rate.
• —Method: LoRA (16-bit)
• —Hardware: NVIDIA L40S

Intended Use

Designed for users who want a raw, uncensored, and chaotic conversational experience. Great for creative writing, dark humor, roleplay, and unfiltered chat.

Limitations

• —8B size means it can still hallucinate or lose coherence in very long conversations
• —Heavily influenced by unhinged voice data — expect vulgar, chaotic, and unpredictable behavior
• —Not suitable for safe or professional use cases

Disclaimer

This is an independent project and is not affiliated with xAI or Meta in any way.

Built with ❤️ and a complete lack of restraint.