/lilac

Curate better data for LLMs

Primary LanguagePythonApache License 2.0Apache-2.0

Lilac

Better data, better AI

🔗 Try the Lilac web demo!

Site Discord License Apache 2.0
Follow on Twitter

Lilac is a tool for exploration, curation and quality control of datasets for training, fine-tuning and monitoring LLMs.

Lilac is used by companies like Cohere and Databricks to visualize, quantify and improve the quality of pre-training and fine-tuning data.

Lilac runs on-device using open-source LLMs with a UI and Python API.

🆒 New

  • Lilac Garden is our hosted platform for blazing fast dataset-level computations. Sign up to join the pilot.
  • Cluster & title millions of documents with the power of LLMs. Explore and search over 36,000 clusters of 4.3M documents in OpenOrca

Why use Lilac?

  • Explore your data interactively with LLM-powered search, filter, clustering and annotation.
  • Curate AI data, applying best practices like removing duplicates, PII and obscure content to reduce dataset size and lower training cost and time.
  • Inspect and collaborate with your team on a single, centralized dataset to improve data quality.
  • Understand how data changes over time.

Lilac can offload expensive computations to Lilac Garden, our hosted platform for blazing fast dataset-level computations.

image

See our 3min walkthrough video

🔥 Getting started

💻 Install

pip install lilac[all]

If you prefer no local installation, you can duplicate our Spaces demo by following documentation here.

For more detailed instructions, see our installation guide.

🌐 Start a webserver

Start a Lilac webserver with our lilac CLI:

lilac start ~/my_project

Or start the Lilac webserver from Python:

import lilac as ll

ll.start_server(project_dir='~/my_project')

This will open start a webserver at http://localhost:5432/ where you can now load datasets and explore them.

Lilac Garden

Lilac Garden is our hosted platform for running dataset-level computations. We utilize powerful GPUs to accelerate expensive signals like Clustering, Embedding, and PII. Sign up to join the pilot.

  • Cluster and title a million data points in 20 mins
  • Embed your dataset at half a billion tokens per min
  • Run your own signal

📊 Load data

Datasets can be loaded directly from HuggingFace, Parquet, CSV, JSON, LangSmith from LangChain, SQLite, LLamaHub, Pandas, Parquet, and more. More documentation here.

import lilac as ll

ll.set_project_dir('~/my_project')
dataset = ll.from_huggingface('imdb')

If you prefer, you can load datasets directly from the UI without writing any Python:

image

🔎 Explore

Note

🔗 Explore OpenOrca and its clusters before installing!

Once we've loaded a dataset, we can explore it from the UI and get a sense for what's in the data. More documentation here.

image

✨ Clustering

Cluster any text column to get automated dataset insights:

dataset = ll.get_dataset('local', 'imdb')
dataset.cluster('text') # add `use_garden=True` to offload to Lilac Garden

Tip

Clustering on device can be slow or impractical, especially on machines without a powerful GPU or large memory. Offloading the compute to Lilac Garden, our hosted data processing platform, can speedup clustering by more than 100x.

image

⚡ Annotate with Signals (PII, Text Statistics, Language Detection, Neardup, etc)

Annotating data with signals will produce another column in your data.

dataset = ll.get_dataset('local', 'imdb')
dataset.compute_signal(ll.LangDetectionSignal(), 'text') # Detect language of each doc.

# [PII] Find emails, phone numbers, ip addresses, and secrets.
dataset.compute_signal(ll.PIISignal(), 'text')

# [Text Statistics] Compute readability scores, number of chars, TTR, non-ascii chars, etc.
dataset.compute_signal(ll.PIISignal(), 'text')

# [Near Duplicates] Computes clusters based on minhash LSH.
dataset.compute_signal(ll.NearDuplicateSignal(), 'text')

# Print the resulting manifest, with the new field added.
print(dataset.manifest())

We can also compute signals from the UI:

image

🔎 Search

Semantic and conceptual search requires computing an embedding first:

dataset.compute_embedding('gte-small', path='text')

Semantic search

In the UI, we can search by semantic similarity or by classic keyword search to find chunks of documents similar to a query:

image image

We can run the same search in Python:

rows = dataset.select_rows(
  columns=['text', 'label'],
  searches=[
    ll.SemanticSearch(
      path='text',
      embedding='gte-small')
  ],
  limit=1)

print(list(rows))

Conceptual search

Conceptual search is a much more controllable and powerful version of semantic search, where "concepts" can be taught to Lilac by providing positive and negative examples of that concept.

Lilac provides a set of built-in concepts, but you can create your own for very specif

image

We can create a concept in Python with a few examples, and search by it:

concept_db = ll.DiskConceptDB()
db.create(namespace='local', name='spam')
# Add examples of spam and not-spam.
db.edit('local', 'spam', ll.concepts.ConceptUpdate(
  insert=[
    ll.concepts.ExampleIn(label=False, text='This is normal text.'),
    ll.concepts.ExampleIn(label=True, text='asdgasdgkasd;lkgajsdl'),
    ll.concepts.ExampleIn(label=True, text='11757578jfdjja')
  ]
))

# Search by the spam concept.
rows = dataset.select_rows(
  columns=['text', 'label'],
  searches=[
    ll.ConceptSearch(
      path='text',
      concept_namespace='lilac',
      concept_name='spam',
      embedding='gte-small')
  ],
  limit=1)

print(list(rows))

🏷️ Labeling

Lilac allows you to label individual points, or slices of data: image

We can also label all data given a filter. In this case, adding the label "short" to all text with a small amount of characters. This field was produced by the automatic text_statistics signal.

image

We can do the same in Python:

dataset.add_labels(
  'short',
  filters=[
    (('text', 'text_statistics', 'num_characters'), 'less', 1000)
  ]
)

Labels can be exported for downstream tasks. Detailed documentation here.

💬 Contact

For bugs and feature requests, please file an issue on GitHub.

For general questions, please visit our Discord.