MFANN-Llama3.1-Abliterated-Slerp-TIES
| Property | Value |
|---|---|
| Parameter Count | 8.03B |
| Model Type | Text Generation |
| Architecture | LLaMA-based Transformer |
| Tensor Type | FP16 |
What is MFANN-Llama3.1-Abliterated-Slerp-TIES?
MFANN-Llama3.1-Abliterated-Slerp-TIES is an advanced language model built on the LLaMA 3.1 architecture, created by netcat420. This model represents a sophisticated merge of multiple base models using MergeKit technology, optimized for both general text generation and specific instruction-following tasks.
Implementation Details
The model implements a context length of 8192 tokens and utilizes specialized sampling parameters including temperature (1.0), top-p (1.0), and top-k (50) settings. It features two distinct operation modes: a standard helpful assistant mode and an experimental "SATANN" mode for cybersecurity-related tasks.
- FP16 precision for efficient computation
- 8192 token context window
- Specialized repeat penalty system (1.19 over 69 tokens)
- GPU layer optimization for Vulkan offloading
Core Capabilities
- Strong performance on IFEval (42.93% accuracy)
- Competitive BBH performance (27.6% normalized accuracy)
- MMLU-PRO capabilities (28.13% accuracy)
- Mathematical reasoning (5.97% on MATH Lvl 5)
- Specialized instruction following and conversational abilities
Frequently Asked Questions
Q: What makes this model unique?
This model stands out for its dual-mode operation capability and its strong performance on instruction-following tasks, particularly achieving 42.93% on IFEval. It uses an innovative merge of multiple LLaMA-based models with SLERP technology.
Q: What are the recommended use cases?
The model is well-suited for text generation tasks, conversation, and instruction following. It offers both a standard helpful assistant mode for general use and a specialized mode for cybersecurity-related applications.
