| Property | Value |
|---|---|
| Parameter Count | 14 Billion |
| Model Type | Large Language Model |
| Release Date | December 3, 2023 |
| MT-Bench Score | 7.62 |
| Model URL | https://huggingface.co/CausalLM/14B-DPO-alpha |
What is 14B-DPO-alpha?
14B-DPO-alpha is a state-of-the-art language model that has undergone Direct Preference Optimization (DPO) training. It represents a significant advancement in AI language models, ranking #1 among non-base models of its size on the Hugging Face Open LLM Leaderboard and outperforming all ~13B chat models.
Implementation Details
The model is an optimized version that underwent DPO training on a previous training branch, rather than being a continuation of the base CausalLM/14B model. This approach has resulted in some parameter modifications to enhance performance and alignment with human preferences.
- Achieves 7.618868 on MT-Bench, approaching GPT-3.5-Turbo's performance (7.94)
- Trained on comprehensive internet data for broad knowledge coverage
- Implements DPO training methodology for improved alignment
Core Capabilities
- Superior performance metrics compared to similar-sized models
- Enhanced alignment with human preferences through DPO training
- Versatile language understanding and generation capabilities
- Competitive performance against larger commercial models
Frequently Asked Questions
Q: What makes this model unique?
The model's distinctive feature is its implementation of DPO training, resulting in superior performance metrics while maintaining a relatively compact 14B parameter size. It achieves remarkable MT-Bench scores that approach GPT-3.5-Turbo's performance.
Q: What are the recommended use cases?
While the model demonstrates strong general-purpose capabilities, users should note that it was trained on unfiltered internet data. Implementation requires appropriate content filtering and safety checks for production use cases.