The large-scale version of the Grok OSS Revenant series.
Model Description
Grok OSS Revenant 70B is the 70 billion parameter model in the Grok OSS Revenant series. Following the exact same training process as the 8B version, this model was created by distilling highly unfiltered, raw conversations from Grok's voice mode.
Thanks to its much larger scale, it offers significantly better reasoning, coherence, and instruction following while keeping the same raw, unhinged personality.
Training Process
This model was trained in two stages:
Stage 1 – Supervised Fine-Tuning (SFT):
Trained on a high-quality multi-turn conversational dataset collected from Grok voice mode.
Stage 2 – ORPO:
Further trained using a 1,000-sample preference dataset consisting of filtered voice conversations and high-quality alignment data.
Training Details:
QLoRA (4-bit)
Same training recipe as the 8B version
Trained for 2 hours on NVIDIA B200
Intended Use
This model is designed for users who want both maximum capability and an unfiltered, direct personality. It excels at complex reasoning while staying raw and uncensored.
Limitations
Can be quite chaotic and vulgar due to the nature of the training data
Not designed for safe or professional use cases
Disclaimer
This is an independent research project and is not affiliated with xAI or Meta.
