Dolphin 2.9 Llama3 8B GGUF
| Property | Value |
|---|---|
| Parameter Count | 8.03B |
| License | META LLAMA 3 COMMUNITY LICENSE |
| Base Model | Meta-Llama-3-8B |
| Context Length | 8K (base model), 4K (fine-tuning) |
| Training Duration | 2.5 days on 8x L40S |
What is dolphin-2.9-llama3-8b-gguf?
Dolphin 2.9 is an advanced language model built on Meta's Llama 3 architecture, developed by Eric Hartford, Lucas Atkins, and Fernando Fernandes at Cognitive Computations. This uncensored model represents a significant advancement in conversational AI, trained using full-parameter fine-tuning with ChatML prompt formatting.
Implementation Details
The model underwent comprehensive training using 10 diverse datasets, including OpenHermes-2.5, CodeFeedback, and UltraChat-200k. It employs the ChatML template format for interactions and was trained using the Axolotl framework with specific optimizations like gradient checkpointing and flash attention.
- Full-weight fine-tuning with 4K sequence length
- Trained using adamw_8bit optimizer with cosine learning rate scheduler
- Implements gradient checkpointing and flash attention for optimal performance
- Uses ChatML prompt template for standardized interactions
Core Capabilities
- Advanced conversational abilities with uncensored responses
- Robust coding and programming support
- Function calling capabilities
- Initial agentic abilities
- Instruction-following and task completion
- Mathematical problem-solving capabilities
Frequently Asked Questions
Q: What makes this model unique?
This model stands out for its uncensored nature and comprehensive training across multiple domains. It combines conversational abilities with coding expertise and function-calling capabilities, all while maintaining compliance with user requests.
Q: What are the recommended use cases?
The model is suitable for programming assistance, conversational applications, and general task automation. However, users should implement their own alignment layer before deploying it as a service, given its uncensored nature.
