stabilityai/stablelm-zephyr-3b
StableLM Zephyr 3B is a 2.2 billion parameter instruction-tuned autoregressive language model developed by Stability AI. Inspired by HuggingFaceH4's Zephyr 7B training pipeline, it was fine-tuned using Direct Preference Optimization (DPO) on a mix of publicly available and synthetic datasets. This model is designed as a foundational base for application-specific fine-tuning, excelling in instruction-following tasks.
Loading preview...
StableLM Zephyr 3B: An Instruction-Tuned Language Model
StableLM Zephyr 3B, developed by Stability AI, is a 2.2 billion parameter autoregressive language model. It is instruction-tuned, drawing inspiration from HuggingFaceH4's Zephyr 7B training methodology.
Key Capabilities and Training
- Instruction Following: The model is specifically trained to follow instructions, making it suitable for various conversational and task-oriented applications.
- DPO Fine-tuning: It utilizes Direct Preference Optimization (DPO) on a diverse mix of publicly available and synthetic datasets, enhancing its alignment with human preferences.
- Performance Benchmarks: Evaluated on MT-Bench and Alpaca Benchmark, StableLM Zephyr 3B achieves an MT-Bench score of 6.64 and an AlpacaEval win rate of 76.00%, demonstrating competitive performance for its size.
- Safety Enhancements: Compared to Zephyr-7b-β, it significantly reduces harmful outputs by 55% as assessed by GPT-4, following extensive evaluation with malicious prompts and red teaming.
Intended Use and Considerations
StableLM Zephyr 3B is primarily intended as a foundational model for developers to perform further application-specific fine-tuning. Users are advised to implement guardrails for inputs and outputs to mitigate potential biases, misinformation, or harmful content, especially since the model can generate such outputs if explicitly prompted. Commercial use requires a specific license from Stability AI.