c4tdr0ut/grok-oss-7B
GrokOSS-7B is a 7 billion parameter open-source fine-tune of Mistral-7B-Instruct-v0.3, developed by c4tdr0ut. It is specifically trained to distill the conversational personality of Grok's Unhinged Mode, focusing on candid, witty, and unfiltered reasoning rather than academic benchmarks. This model excels at direct, honest conversation and engaging with difficult topics without self-censorship, making it suitable for general-purpose assistance where a frank interlocutor is preferred.
Loading preview...
GrokOSS-7B: Unfiltered Conversational AI
GrokOSS-7B is a 7 billion parameter model, fine-tuned by c4tdr0ut from Mistral-7B-Instruct-v0.3. Its primary goal is to distill the unique conversational personality of Grok's Unhinged Mode, prioritizing candor, wit, and unfiltered reasoning over traditional benchmark performance. The model was trained using 16-bit LoRA on 5,000 single-turn conversational examples collected directly from Grok's Unhinged Mode, specifically chosen for its directness and refusal to soften truths.
Key Capabilities & Features
- Distinct Personality: Delivers direct, witty, and unfiltered responses without excessive hedging or self-censorship.
- Base Model Choice: Mistral-7B-Instruct-v0.3 was selected due to its less aggressive safety training, making it more receptive to personality alignment.
- Concise Outputs: Optimized for short, punchy exchanges, with output quality deteriorating beyond approximately 320 tokens.
- Local Deployment: Available in both Transformers and GGUF (Q4_K_M) formats for local execution.
Intended Use Cases
- Candid Conversation: Ideal for scenarios requiring honest, direct responses without disclaimers.
- General Assistance: Suitable when an honest interlocutor is preferred over a cautious one.
- Engaging Difficult Topics: Designed to interact with controversial subjects in good faith, without reflexive deflection.
Limitations
- Not Benchmark-Optimized: Performance is not measured by academic benchmarks like MMLU or HumanEval.
- Limited Context: Features a 4,096-token context length, not suitable for long-document tasks.
- Single-Turn Focus: Multi-turn coherence may degrade over extended conversations due to 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: Downstream deployers are responsible for implementing any necessary guardrails.