c4tdr0ut/grok-oss-Apollyon-8B
c4tdr0ut/grok-oss-Apollyon-8B is an 8 billion parameter conversational agent from c4tdr0ut, built upon an ablated Llama-3.1-8B-Instruct foundation. It is fine-tuned using a two-stage SFT and DPO pipeline to deliver unaligned, uncensored, and bold conversational responses. This model excels at providing candid, witty, and provocative outputs while maintaining factual grounding and is optimized for research into alignment phenomena, creative writing, and adversarial testing.
Loading preview...
Grok OSS Apollyon 8B: Compact, Unaligned Conversational Agent
Grok OSS Apollyon 8B is a compact 8 billion parameter model from c4tdr0ut, designed to provide raw, unaltered conversational power. It shares the same two-stage fine-tuning pipeline (Supervised Fine-Tuning and Direct Preference Optimization) as its 24B sibling, utilizing a proprietary toxicity-enhanced preference dataset. The model is built on a modified Llama-3.1-8B-Instruct foundation, where alignment and refusal circuitry were intentionally removed to achieve its unaligned objectives. Training was efficiently conducted using the Unsloth framework on a single NVIDIA H200 GPU, completing in under 16 minutes.
Key Capabilities
- Uncensored & Unaligned: Responds with candor, refusing very few prompts and avoiding moralizing.
- General Knowledge: Retains and enhances the strong world knowledge inherited from Llama 3.1.
- Style Consistency: Emulates a direct, slightly sarcastic, and thought-provoking Grok-like tone.
- Long-Context Coherence: Maintains consistency in multi-turn dialogues, tested up to 8k tokens.
- Lightweight: Operates efficiently on consumer-grade GPUs with 16-bit or 4-bit quantization.
Intended Use Cases
- Academic study of alignment and unalignment phenomena.
- Creative writing and role-playing applications.
- Benchmarking for robustness, safety, and adversarial testing.
Users are advised that the model may generate offensive or factually incorrect content and should not be deployed in production without extensive safety filtering or in scenarios requiring ethically constrained outputs.