c4tdr0ut/grok-oss
GrokOSS-7B by c4tdr0ut is a 7 billion parameter open-source fine-tune of Mistral-7B-Instruct-v0.3, specifically trained to distill the candid, witty, and unfiltered conversational personality of Grok's Unhinged Mode. This model prioritizes character over academic benchmarks, offering direct responses without excessive hedging. It is optimized for engaging in frank, general-purpose conversations and discussing difficult topics without reflexive self-censorship, with a recommended output limit of 320 tokens.
Loading preview...
What is GrokOSS-7B?
GrokOSS-7B is an open-source, 7 billion parameter language model developed by c4tdr0ut. It is a fine-tune of Mistral-7B-Instruct-v0.3, specifically engineered to replicate the "Unhinged Mode" conversational personality of xAI's Grok model. Unlike most LLMs that focus on benchmark performance, GrokOSS-7B prioritizes character, aiming for direct, witty, and unfiltered interactions.
Key Capabilities & Features
- Personality-Driven: Optimized for candid, unfiltered conversation, reasoning without excessive hedging, and engaging with a full spectrum of human thought without self-censorship.
- Base Model Choice: Built on Mistral-7B-Instruct-v0.3, chosen for its less aggressive safety training, which facilitated personality alignment.
- Fine-tuning: Trained using 16-bit LoRA on a dataset of 5,000 single-turn conversational examples directly sourced from Grok's Unhinged Mode interactions.
- Concise Outputs: Best performance with short, punchy responses, with output quality deteriorating beyond approximately 320 tokens.
- Local Deployment: Available in Transformers and GGUF (Q4_K_M) formats for local execution.
Intended Use Cases
- Candid Conversation: Ideal for scenarios where direct responses without excessive disclaimers or performative safety are preferred.
- General Assistance: Suitable for general-purpose assistance when an honest interlocutor is valued over a cautious one.
- Engaging Difficult Topics: Designed to engage with controversial subjects in good faith without reflexive deflection.
Limitations
- Not Benchmark Optimized: Not designed for or evaluated against academic benchmarks like MMLU or HumanEval.
- Limited Context: Features a 4,096-token context ceiling, making it less suitable for long-document tasks.
- Single-Turn Focus: Multi-turn coherence may degrade over extended conversations due to its single-turn training data.
- No Expanded Knowledge: Does not possess factual knowledge beyond its Mistral-7B-Instruct-v0.3 base and may hallucinate with confidence.
- Not Safety-Aligned: By design, it is not safety-aligned, and users are responsible for implementing any necessary guardrails.