Generalist large language models can be improved for specific use cases by fine-tuning them with the open-source project InstructLab on consumer hardware.
Generalist LLMs often fail on specific or recent queries due to knowledge cutoffs, demonstrating the need for adaptation for real-world use cases.
Generative AI presents significant challenges including legal exposure, hallucinations, hiring bias, and high operational costs for tuning and inference.
Models can be improved through methods like alignment tuning, which specializes the model's core knowledge, unlike RAG which only supplements it with external context.
Selecting a smaller, fine-tuned foundation model like Granite can drastically reduce operational costs compared to using a large, general-purpose model.
InstructLab uses a simple YAML taxonomy to automatically generate large synthetic training datasets, making model tuning accessible to non-data scientists.
The InstructLab CLI provides a streamlined workflow for initializing a project, downloading a model, generating synthetic data, and training the model locally.
A pre-trained model can be enhanced with specific enterprise domain knowledge, such as insurance claim data, using the InstructLab tuning process.
The open-source InstructLab project offers community resources like GitHub, Slack, and a mailing list for developers to get involved.
35:16Ankit Patel
58:00Vijay Krishan Gupta & Gauravdeep Singh Lotey
34:21Aarno Aukia
29:11Michael Hunger
23:50Aditi Godbole
29:40Jodie Burchell
30:04Roberto Carratalá & Cedric Clyburn
27:11Emanuele Fabbiani
.png?w=240&auto=compress,format)
.gif?w=240&auto=compress,format)
Jobs that call for the skills explored in this talk.
LEONHARD WEISS GmbH & Co. KG
Red Hat