IBM has released the artificial intelligence model Granite, which was trained on a database of code in 116 languages. The model comes in 4 sizes: 3B, 8B, 20B, and 34B, with base and instructable versions. Particularly, the 8B model excels in programming capabilities compared to other models of the same size.
The training process of the 34B version is unique as it utilizes upscaling by extracting 8 layers of output from the 20B model, which has 52 layers, and connecting it with another 8-layer input model to form the 34B. The training data includes GitHub Code Clean, StarCoderData, and other code sources.
All models are available under the Apache 2.0 license, allowing for more flexibility in usage compared to previous models. IBM has stated that they are working on other models as well and will release them in the future.
TLDR: IBM has launched the advanced AI model Granite trained on a diverse dataset of code in multiple languages, with different sizes and capabilities, offering flexibility in usage.
Leave a Comment