Foundation Model Fine-tuning

Preview

This feature is in Public Preview in us-east-1 and us-west-2.

With Foundation Model Fine-tuning (now part of Mosaic AI Model Training), you can use your own data to customize a foundation model to optimize its performance for your specific application. By conducting full parameter fine-tuning or continuing training of a foundation model, you can train your own model using significantly less data, time, and compute resources than training a model from scratch.

With Databricks you have everything in a single platform: your own data to use for training, the foundation model to train, checkpoints saved to MLflow, and the model registered in Unity Catalog and ready to deploy.

See Tutorial: Create and deploy a Foundation Model Fine-tuning run to learn how to create a run using the Foundation Model Fine-tuning API, and then review the results and deploy the model using the Databricks UI and Mosaic AI Model Serving.

What is Foundation Model Fine-tuning?

Foundation Model Fine-tuning lets you use the Databricks API or UI to tune or further train a foundation model.

Using Foundation Model Fine-tuning, you can:

Train a model with your custom data, with the checkpoints saved to MLflow. You retain complete control of the trained model.
Automatically register the model to Unity Catalog, allowing easy deployment with model serving.
Further train a completed, proprietary model by loading the weights of a previously trained model.

Databricks recommends that you try Foundation Model Fine-tuning if:

You have tried few-shot learning and want better results.
You have tried prompt engineering on an existing model and want better results.
You want full ownership over a custom model for data privacy.
You are latency-sensitive or cost-sensitive and want to use a smaller, cheaper model with your task-specific data.

Supported tasks

Foundation Model Fine-tuning supports the following use cases:

Chat completion: Recommended task. Train your model on chat logs between a user and an AI assistant. This format can be used both for actual chat logs, and as a standard format for question answering and conversational text. The text is automatically formatted into the appropriate format for the specific model. See example chat templates in the HuggingFace documentation for more information on templating.
Supervised fine-tuning: Train your model on structured prompt-response data. Use this to adapt your model to a new task, change its response style, or add instruction-following capabilities. This task does not automatically apply any formatting to your data and is only recommended when custom data formatting is required.
Continued pre-training: Train your model with additional text data. Use this to add new knowledge to a model or focus a model on a specific domain.

Requirements

A Databricks workspace in one of the following AWS regions: us-east-1 and us-west-2.
Foundation Model Fine-tuning APIs installed using pip install databricks_genai.
Your workspace must not use S3 access policies.
Databricks Runtime 12.2 LTS ML or above if your data is in a Delta table.

See Prepare data for Foundation Model Fine-tuning for information about required input data formats.

Recommended data size for model training

Databricks recommends initially training using one to four epochs. After evaluating your fine-tuned model, if you want the model outputs to be more similar to your training data, you can start to continue training by using one to two more epochs.

If the model performance significantly decreases on tasks not represented in your fine-tuning data, or if the model appears to output exact copies of your fine-tuning data, Databricks recommends reducing the number of training epochs.

For supervised fine-tuning and chat completion, you should provide enough tokens for at least one full context length of the model. For example, 4096 tokens for meta-llama/Llama-2-7b-chat-hf or 32768 tokens for mistralai/Mistral-7B-v0.1.

For continued pre-training, Databricks recommends a minimum of 1.5 million tokens to get a higher quality model that learns your custom data.

Supported models

The following table lists supported models. See Model licenses for the applicable model license and acceptable use policy information.

To continue support of the most state-of-the-art models, Databricks might update supported models or retire older models. See To be retired models.

Model	Maximum context length	Notes
`databricks/dbrx-base`	32768
`databricks/dbrx-instruct`	32768
`meta-llama/Llama-3.2-1B`	131072
`meta-llama/Llama-3.2-1B-Instruct`	131072
`meta-llama/Llama-3.2-3B`	131072
`meta-llama/Llama-3.2-3B-Instruct`	131072
`meta-llama/Meta-Llama-3.1-405B`	131072
`meta-llama/Meta-Llama-3.1-405B-Instruct`	131072
`meta-llama/Meta-Llama-3.1-70B`	131072
`meta-llama/Meta-Llama-3.1-70B-Instruct`	131072
`meta-llama/Meta-Llama-3.1-8B`	131072
`meta-llama/Meta-Llama-3.1-8B-Instruct`	131072
`mistralai/Mistral-7B-v0.1`	32768
`mistralai/Mistral-7B-Instruct-v0.2`	32768
`mistralai/Mixtral-8x7B-v0.1`	32768

To be retired models

The following table lists supported models that are planned for retirement. See Retired models for planned retirement dates and recommended model replacements.

Model	Maximum context length	Notes
`meta-llama/Meta-Llama-3-70B`	8192	This model is no longer supported after January 7, 2025.
`meta-llama/Meta-Llama-3-70B-Instruct`	8192	This model is no longer supported after January 7, 2025.
`meta-llama/Meta-Llama-3-8B`	8192	This model is no longer supported after January 7, 2025.
`meta-llama/Meta-Llama-3-8B-Instruct`	8192	This model is no longer supported after January 7, 2025.
`meta-llama/Llama-2-7b-hf`	4096	This model is no longer supported after January 7, 2025.
`meta-llama/Llama-2-13b-hf`	4096	This model is no longer supported after January 7, 2025.
`meta-llama/Llama-2-70b-hf`	4096	This model is no longer supported after January 7, 2025.
`meta-llama/Llama-2-7b-chat-hf`	4096	This model is no longer supported after January 7, 2025.
`meta-llama/Llama-2-13b-chat-hf`	4096	This model is no longer supported after January 7, 2025.
`meta-llama/Llama-2-70b-chat-hf`	4096	This model is no longer supported after January 7, 2025.
`codellama/CodeLlama-7b-hf`	16384	This model is no longer supported after January 7, 2025.
`codellama/CodeLlama-13b-hf`	16384	This model is no longer supported after January 7, 2025.
`codellama/CodeLlama-34b-hf`	16384	This model is no longer supported after January 7, 2025.
`codellama/CodeLlama-7b-Instruct-hf`	16384	This model is no longer supported after January 7, 2025.
`codellama/CodeLlama-13b-Instruct-hf`	16384	This model is no longer supported after January 7, 2025.
`codellama/CodeLlama-34b-Instruct-hf`	16384	This model is no longer supported after January 7, 2025.
`codellama/CodeLlama-7b-Python-hf`	16384	This model is no longer supported after January 7, 2025.
`codellama/CodeLlama-13b-Python-hf`	16384	This model is no longer supported after January 7, 2025.
`codellama/CodeLlama-34b-Python-hf`	16384	This model is no longer supported after January 7, 2025.

Model licenses

The following table provides the applicable model license and acceptable use policy information for the supported model families.

Model family	License and acceptable use policy
Meta Llama 3.2	Meta Llama 3.2 is licensed under the LLAMA 3.2 Community License, Copyright © Meta Platforms, Inc. All Rights Reserved. Customers are responsible for ensuring their compliance with the terms of this license and the Llama 3.2 Acceptable Use Policy.
Meta Llama 3.1	Meta Llama 3.1 is licensed under the LLAMA 3.1 Community License, Copyright © Meta Platforms, Inc. All Rights Reserved. Customers are responsible for ensuring compliance with applicable model licenses.
Llama 3	Llama 3 is licensed under the LLAMA 3 Community License, Copyright © Meta Platforms, Inc. All Rights Reserved. Customers are responsible for ensuring compliance with applicable model licenses.
Llama 2	Llama 2 is licensed under the LLAMA 2 Community License, Copyright © Meta Platforms, Inc. All Rights Reserved. Customers are responsible for ensuring compliance with applicable model licenses.
CodeLlama	CodeLlama models are licensed under the LLAMA 2 Community License, Copyright © Meta Platforms, Inc. All Rights Reserved. Customers are responsible for ensuring compliance with applicable model licenses.
DBRX	DBRX is provided under and subject to the Databricks Open Model License, Copyright © Databricks, Inc. All rights reserved. Customers are responsible for ensuring compliance with applicable model licenses, including the Databricks Acceptable Use policy.

Use Foundation Model Fine-tuning

Foundation Model Fine-tuning is accessible using the databricks_genai SDK. The following example creates and launches a training run that uses data from Unity Catalog Volumes. See Create a training run using the Foundation Model Fine-tuning API for configuration details.

from databricks.model_training import foundation_model as fm

model = 'meta-llama/Meta-Llama-3.1-8B-Instruct'
# UC Volume with JSONL formatted data
train_data_path = 'dbfs:/Volumes/main/mydirectory/ift/train.jsonl'
register_to = 'main.mydirectory'
run = fm.create(
  model=model,
  train_data_path=train_data_path,
  register_to=register_to,
)

See the Instruction fine-tuning: Named Entity Recognition demo notebook for an instruction fine-tuning example that walks through data preparation, fine-tuning training run configuration and deployment.

Limitations

Large datasets (10B+ tokens) are not supported due to compute availability.
For continuous pre-training, workloads are limited to 60-256MB files. Files larger than 1GB may cause longer processing times.
Databricks strives to make the latest state-of-the-art models available for customization using Foundation Model Fine-tuning. As new models become available, access to older models from the API or UI might be removed, older models might be deprecated, or supported models updated. See Generative AI models maintenance policy.