Phi-2-super
| Property | Value |
|---|---|
| Parameter Count | 2.78B |
| License | MIT |
| Tensor Type | BF16 |
| Language | English |
What is phi-2-super?
Phi-2-super is an enhanced version of Microsoft's Phi-2 model, fine-tuned using SFT (Supervised Fine-Tuning) and cDPO techniques. This 2.78B parameter model is specifically designed for improved instruction following and conversational abilities, achieving a prompt-level loose accuracy of 0.2717 on the Instruction Following Eval dataset.
Implementation Details
The model implements a chat template similar to Mistral instruct models, utilizing the transformers library for inference. It supports efficient text generation with customizable parameters like temperature and top-p sampling.
- Supports batch processing and GPU acceleration
- Uses BF16 tensor format for efficient computation
- Implements sophisticated chat templating for conversation handling
- Provides built-in safety measures and token management
Core Capabilities
- Text generation and completion tasks
- Conversational AI applications
- Instruction following with high accuracy
- Code generation and technical writing
- Custom inference endpoint support
Frequently Asked Questions
Q: What makes this model unique?
The model combines the efficiency of the Phi-2 architecture with enhanced instruction-following capabilities through SFT and cDPO optimization, making it particularly suitable for conversational AI applications while maintaining a relatively small parameter count.
Q: What are the recommended use cases?
The model excels in conversational AI, code generation, and general text generation tasks. It's particularly well-suited for applications requiring both computational efficiency and high-quality instruction following capabilities.
