c4tdr0ut/grok-oss-Apollyon-24B
Grok OSS Apollyon 24B by c4tdr0ut is a 24 billion parameter unaligned conversational agent built on the Mistral-Small-24B-Instruct-2501 foundation. Fine-tuned using a two-stage SFT and DPO pipeline, it excels at generating high-fidelity text with superior general knowledge, sharp wit, and a distinctly "unhinged" persona. This model is primarily intended for research into alignment phenomena, creative writing, and adversarial testing, offering uncensored and candid responses.
Loading preview...
Grok OSS Apollyon 24B: An Unaligned Conversational Agent
Grok OSS Apollyon 24B is the fourth model in the Grok OSS series, developed by c4tdr0ut to explore the boundaries of unaligned, high-fidelity text generation. Built upon an "abliterated" Mistral-Small-24B-Instruct-2501 foundation, this 24 billion parameter model underwent a two-stage fine-tuning process: Supervised Fine-Tuning (SFT) on a proprietary dataset of ~1,000 conversational turns, followed by Direct Preference Optimization (DPO) using a "toxicity-enhanced" preference dataset. This methodology aimed to produce a model with superior general knowledge, sharp wit, and a distinctly "unhinged" persona, while maintaining coherence.
Key Capabilities
- Uncensored & Unaligned: Designed to refuse very few prompts, responding with candor and without moralizing.
- General Knowledge: Retains and enhances the base model’s broad world knowledge.
- Style Consistency: Emulates a direct, slightly sarcastic, and thought-provoking "Grok tone."
- Long-Context Coherence: Maintains consistency over multi-turn dialogues, tested up to 8k tokens.
Intended Use & Limitations
This model is a research artifact for academic study of alignment/unalignment, creative writing, role-playing, and adversarial testing. It is not recommended for production deployment without extensive safety filtering due to its potential to generate offensive, biased, or factually incorrect content. Users should be aware of known quirks, including occasional erratic or overly hostile responses due to minor training data contamination, which can often be resolved by clearing context and re-asking.