Dolphin X1 8B: Uncensored, Steerable General-Purpose LLM
Dolphin X1 8B is an 8 billion parameter language model developed by dphn, built upon the Llama 3.1 8B Instruct architecture. Its primary goal is to provide an uncensored version of Llama 3.1 while preserving or enhancing its core abilities through fine-tuning. Unlike many proprietary models, Dolphin X1 8B emphasizes user control, allowing system owners to define system prompts and alignment guidelines, ensuring data privacy and custom ethical frameworks.
Key Capabilities & Differentiators
- Uncensored Foundation: Directly uncensors Llama 3.1 8B Instruct, offering greater flexibility in content generation.
- User Steerability: Provides full control over system prompts and alignment, enabling users to dictate the model's tone, character, and behavioral rules.
- Data Privacy: Designed to keep user queries private, preventing external entities from accessing or utilizing sensitive data.
- General Purpose: Aims to function as a versatile model, similar to commercial offerings like ChatGPT, Claude, and Gemini, but with enhanced user autonomy.
- Consistent Behavior: Mitigates issues faced with proprietary models, such as unexpected changes in system prompts, model versions, or alignment, by giving control back to the user.
Performance Metrics
Evaluations indicate solid performance across various benchmarks:
- MMLU: 0.626900
- IFEVAL: 0.608100
- Dolphin-refusals: 95.96% pass rate on 4.5k commonly refused prompts.
Use Cases
Dolphin X1 8B is ideal for businesses and developers who require:
- Customizable AI Applications: Where specific ethical guidelines and content moderation are managed internally.
- Data-Sensitive Environments: Applications where query data must remain private and under user control.
- Stable AI Integration: Projects that need consistent model behavior and versioning without external interference.
- Flexible Content Generation: Scenarios benefiting from an uncensored model that can be aligned to specific use cases.