Medius-Erebus-Magnum-14B-GGUF
| Property | Value |
|---|---|
| Parameter Count | 14.8B |
| Model Type | Transformers/GGUF |
| Base Model | Qwen/Qwen2.5-14B |
| Training Framework | Axolotl 0.4.1 |
What is medius-erebus-magnum-14b-GGUF?
Medius-Erebus-Magnum is a quantized version of the original model created using llama.cpp, built on the foundation of Qwen2.5-14B. It represents a significant advancement in conversational AI, leveraging advanced training techniques and optimization strategies.
Implementation Details
The model was trained using state-of-the-art techniques including gradient checkpointing with unsloth, flash attention, and specialized plugins like LigerPlugin. Training utilized a sequence length of 32768 with sample packing, and employed the AdamW 8-bit optimizer with a cosine learning rate scheduler.
- Training conducted over 2 epochs with a learning rate of 0.000008
- Implements flash attention and RoPE optimizations
- Uses ChatML template for conversation formatting
- Trained on diverse datasets including instruction and conversation data
Core Capabilities
- Advanced conversational abilities with extended context handling
- Optimized for efficiency through GGUF quantization
- Enhanced attention mechanisms through Liger optimizations
- Robust instruction following and response generation
Frequently Asked Questions
Q: What makes this model unique?
The model combines the powerful Qwen2.5-14B architecture with advanced optimization techniques, including GGUF quantization and specialized attention mechanisms, making it both powerful and efficient for deployment.
Q: What are the recommended use cases?
This model is particularly well-suited for conversational applications, instruction-following tasks, and general-purpose language understanding and generation where efficiency and performance are crucial.
