Overview

This model, wandb/gemma-7b-zephyr-sft, is an 8.5 billion parameter GPT-like language model. It is a supervised fine-tuned (SFT) version of Google's gemma-7b model, utilizing the Zephyr SFT recipe from the Hugging Face alignment handbook. The training process was logged and can be explored in the associated Weights & Biases workspace.

Key Capabilities & Performance

The model is primarily designed for English language tasks and demonstrates competitive performance on various benchmarks, as evaluated on the Open LLM Leaderboard. Key scores include:

• Avg. Score: 61.64
• HellaSwag (10-Shot): 80.73
• MMLU (5-Shot): 60.33
• Winogrande (5-shot): 74.19
• GSM8k (5-shot): 49.81

Use Cases

Given its fine-tuned nature and benchmark performance, this model is well-suited for general instruction-following tasks, conversational AI, and applications requiring strong reasoning and common sense understanding. Its foundation on Gemma 7B combined with the Zephyr SFT approach aims to provide a robust base for various NLP applications.