Try the Lilac hosted demo on HuggingFace or find us on GitHub: github.com/lilacai/lilac

Introducing Lilac#

August 21, 2023
Daniel Smilkov & Nikhil Thorat

Lilac is an open-source tool that enables AI practitioners to see and quantify their datasets.

For an end-to-end example, see our Quick Start guide.

For a detailed documentation, visit our website.

Lilac allows users to:

  • Browse datasets with unstructured data.

  • Enrich unstructured fields with structured metadata using Lilac Signals, for instance near-duplicate and personal information detection. Structured metadata allows us to compute statistics, find problematic slices, and eventually measure changes over time.

  • Create and refine Lilac Concepts which are customizable AI models that can be used to find and score text that matches a concept you may have in your mind.

  • Download the results of the enrichment for downstream applications.

Out of the box, Lilac comes with a set of generally useful Signals and Concepts, however this list is not exhaustive and we will continue to work with the OSS community to continue to add more useful enrichments.

Our mission#

At Lilac, our mission is to make unstructured data visible, quantifiable, and malleable.

This will lead to:

  • Higher quality AI models

  • Better actionability when AI models fail

  • Better control and visibility of model bias

Data quality in AI is tricky#

During our time at Google, we collaborated with many teams to improve datasets used to build their AI models. Their goal was to continually improve the quality of their models, often focusing on refining the training data.

What makes improving data quality difficult is that many AI models rely on unstructured data, such as natural language or images, that lack any labels or useful metadata. To complicate matters, what constitutes “good” data depends heavily on the application and the user experience. Despite these differences, a common thread emerged: while teams would compute aggregate statistics to understand the general composition of their data, they often overlooked the raw data. When methodically organized and visualized, glaring bugs in datasets would present themselves, often with simple fixes leading to higher quality models.

“Bad data”#

“Bad data” is often hard to define, but we often know bad data when we see it. In other cases, “bad data” isn’t objectively bad: for instance, the presence of German text in a French to English translation dataset will negatively affect the translation model, even if the translation is correct for German.

With that observation in mind, at Google we built tools and processes that empowered teams to see their data. To summarize a few years of learning into one sentence: each dataset has its own quirks, and these quirks can have non-obvious implications for the quality of downstream models.

Today, data cleaning for datasets fed into AI models is often done with heuristics in a Python script, with little visibility into the side effects of that change.


Since each AI application has its own requirements, we’re focused on enabling users to annotate data with customizable Concepts. Concepts can be created and refined through the UI, and updated in real-time with user-feedback. These AI-powered embedding-based classifiers can be specific to an application, e.g. termination clauses in legal contracts, or generally applicable, e.g. toxicity.


Data privacy is an important consideration for most AI teams, so we are focused on making Lilac fast and usable with data staying on-premise. Lilac Concepts utilize powerful on-device embeddings, like GTE. However if your application is not sensitive to data privacy (e.g. using open-source datasets), you may choose to use more powerful embeddings like OpenAI, Cohere, PaLM, or your own! For more information on embeddings, see our documentation.

HuggingFace demo#

We are also hosting a HuggingFace space with a handful of popular datasets (e.g. OpenOrca) and curated concepts (e.g. profanity, legal termination, source-code detection, etc.). In this demo, you can browse pre-enriched datasets and even create your own concepts. The space can be forked and made private with your own data, skipping the installation process of Lilac.

Open source#

We believe an open-source product is the best way to improve the culture around dataset quality.

We encourage the AI community to try the tool and help us grow a central repository of useful concepts and signals. We would love to collaborate to shed light on the most popular AI datasets.

Let’s visualize, quantify, and ultimately improve all unstructured datasets.