Merlinite-7B-Lab

A 7-billion parameter model fine-tuned using IBM's LAB methodology to align large language models with high-quality synthetic data.

Merlinite-7B-Lab leverages the Large-scale Alignment Baseline (LAB) technique to improve model performance without exhaustive manual labeling. Built on the Mistral-7B base, it utilizes a taxonomy-driven synthetic data generation process to enhance specific skills and knowledge domains. This approach allows the model to achieve competitive benchmarks (scoring high on MT-Bench) while maintaining a smaller footprint. It is a direct result of the collaboration between IBM Research and Red Hat to streamline the generative AI lifecycle.

https://huggingface.co/ibm/merlinite-7b-lab