72B-Qwen2.5-Kunou-v1-FP8-Dynamic
| Property | Value |
|---|---|
| Base Model | Qwen/Qwen2.5-72B-Instruct |
| Format | ChatML |
| Hugging Face | Link |
| Author | Infermatic |
What is 72B-Qwen2.5-Kunou-v1-FP8-Dynamic?
This is a fine-tuned version of the Qwen2.5-72B model, optimized for generalist tasks and roleplay scenarios. The model represents an evolution in Infermatic's series, building upon lessons learned from previous models like L3-70B-Euryale-v2.2. It utilizes a refined and cleaned-up dataset focusing on creative and instructional content.
Implementation Details
The model employs several advanced technical features, including FP8 dynamic quantization, flash attention, and specialized optimizations through the Axolotl framework version 0.5.2. It utilizes LoRA adaptation with specific parameters (r=128, alpha=16) and implements various efficiency improvements like Liger plugin integrations.
- Sequence length: 16384 tokens
- Learning rate: 0.000004 with cosine scheduler
- Gradient accumulation steps: 4
- Optimizer: paged_ademamix_8bit
Core Capabilities
- Generalist task handling
- Roleplay and creative writing
- Extended context processing
- Efficient memory utilization through quantization
- Improved dataset quality for creative domains
Frequently Asked Questions
Q: What makes this model unique?
The model combines high-quality training data with efficient quantization techniques, making it particularly suitable for both general-purpose applications and creative tasks. It builds upon previous successful models while incorporating refined datasets and modern optimization techniques.
Q: What are the recommended use cases?
The model is optimized for generalist tasks and roleplay scenarios, with recommended settings of Temperature: 1.1 and min_p: 0.1. It's particularly well-suited for creative writing and interactive conversations using the ChatML format.
