December 2023 Summaries

The author, Kyle Corbitt, of OpenPipe, a platform for fine-tuning models, announces the release of "Mistral 7B Fine-Tune Optimized", a new model optimized to be the strongest base model for further fine-tunes. This model outperforms other variants in fine-tuning tasks, with two of its variants, Hermes Neural and Metamath Cybertron Starling, being among the best-performing models overall despite not being directly fine-tuned. The author explains that this is due to a phenomenon called model merging, where combining weights from different models can produce stronger results. The new model has been tested on various tasks and datasets, including those of GPT-4, and shows promising results, with one variant slightly outperforming GPT-4 in some cases.

At OpenPipe, a fully-managed fine-tuning platform for developers is now available, allowing users to replace their existing prompts with fine-tuned models in just a few minutes. The platform captures existing prompts and completions, synthesizes them into a dataset, and fine-tunes models that are a drop-in replacement for the prompt. Starting today, OpenPipe also automates the process of evaluating model performance by using GPT-4 to compare output across multiple models on a test set, allowing users to view results, see which model won, and access custom instructions and reasoning. This feature is designed to help users build and improve their fine-tuned models with ease.