What is Databricks Dolly?

Dolly (Dolly 2.0 and later) is an instruction-following Large Language Model developed by Databricks, built on the EleutherAI Pythia model family. It is distinguished by its **truly open-source nature**, including the model weights and the high-quality human-generated instruction dataset (databricks-dolly-15k), all licensed for commercial use. This allows organizations to own, customize, and deploy the model entirely within their private infrastructure, eliminating vendor API reliance and data leakage risks.

Key Features & Capabilities

• Open & Commercial Use: No licensing restrictions for commercial applications, enabling complete model ownership.
• Instruction Tuning: Fine-tuned on a high-quality, human-generated instruction set for superior command following.
• Data Centric: Designed to be fine-tuned and specialized on private, proprietary datasets within the Databricks Lakehouse Platform.
• Code Generation: Excels at generating and analyzing code, particularly PySpark and SQL for data engineering tasks.

How to Deploy and Use Dolly

Dolly is not used as a public API but as a model deployed on private infrastructure:

1. Deployment: Download the Dolly model weights from Hugging Face Hub or access them directly via the Databricks Lakehouse Platform.
2. Compute Setup: Deploy the model onto a dedicated compute cluster (e.g., a GPU-enabled Databricks cluster).
3. RAG Implementation: For Q&A use cases, ingest and clean proprietary Q&A data, transform it into embeddings, and index it in a vector database.
4. Inference: Utilize **LangChain** or custom Python code within a Databricks notebook to fetch context from the vector database and craft a prompt for Dolly (Retrieval Augmented Generation).
5. Fine-Tuning (Optional): If required, fine-tune the base Dolly model on specialized private datasets to customize its output style and knowledge domain.