Zebrafish-7B: A Model Stock Merge

Zebrafish-7B is a 7 billion parameter language model developed by mlabonne, notable for its use of the innovative Model Stock merge method. This model combines the strengths of liminerity/M7-7b and rwitz/experiment26-truthy-iter-0 using LazyMergekit, built on the mistralai/Mistral-7B-v0.1 base.

Key Capabilities & Performance

Zebrafish-7B exhibits robust general-purpose performance, as evidenced by its evaluation on the Nous benchmark. It achieves an average score of 62.41, with specific scores including:

• AGIEval: 44.92
• GPT4All: 77.18
• TruthfulQA: 78.25
• Bigbench: 49.28

These scores position Zebrafish-7B competitively among other 7B models, including mlabonne/AlphaMonarch-7B and mistralai/Mistral-7B-Instruct-v0.2.

When to Use Zebrafish-7B

• General Language Generation: Its balanced performance across various benchmarks makes it suitable for a broad spectrum of text generation tasks.
• Exploration of Merged Models: Developers interested in the Model Stock merging technique can use this model as a practical example.
• Applications requiring a 7B model: For scenarios where a 7B parameter model is optimal for resource constraints while still delivering strong results.