| Property | Value |
|---|---|
| Parameter Count | 8.03B |
| Model Type | Text Generation |
| Architecture | LLaMA-3 Based |
| License | Other |
| Tensor Type | BF16 |
What is Hathor_Stable-v0.2-L3-8B?
Hathor_Stable-v0.2-L3-8B is an advanced language model built upon the LLaMA-3 8B instruct architecture. This model represents a significant enhancement through comprehensive training across three epochs using a diverse dataset including private data, synthetic opus instructions, and a curated mix of light and classical novel data, along with roleplaying chat pairs.
Implementation Details
The model demonstrates impressive performance across various benchmarks, particularly excelling in instruction following with a 71.75% accuracy on IFEval (0-Shot). Its architecture leverages the latest advances in transformer technology while maintaining efficiency with BF16 tensor operations.
- 8.03B parameters optimized for balanced performance
- Trained on multiple data types including synthetic and novel content
- Specialized instruction-following capabilities
- Available in multiple quantized versions
Core Capabilities
- Strong performance on IFEval (71.75% accuracy)
- Decent showing on BBH 3-shot tasks (32.83%)
- MMLU-PRO performance of 29.96% (5-shot)
- Mathematical reasoning capabilities (9.21% on MATH Level 5)
Frequently Asked Questions
Q: What makes this model unique?
The model's unique strength lies in its comprehensive training approach, combining private data with synthetic instructions and novel content, resulting in particularly strong instruction-following capabilities.
Q: What are the recommended use cases?
Given its performance profile, this model is well-suited for conversational AI applications, instruction-following tasks, and general text generation scenarios where balanced performance across multiple domains is required.