🔍

Chroma Datasets

pip install chroma_datasets

• a public package registry of sample and useful datasets to use with embeddings
• a set of tools to export and import Chroma collections

We built to enable faster experimentation: There is no good source of sample datasets and sample datasets are incredibly important to enable fast experiments and learning.

Current Datasets

Dataset	Size	Contributor	Python Class
State of the Union	51kb	Chroma	from chroma_datasets import StateOfTheUnion
Paul Graham Essay	1.3mb	Chroma	from chroma_datasets import PaulGrahamEssay
Huberman Podcasts	4.3mb	Dexa AI	from chroma_datasets import HubermanPodcasts
more soon...			read below how to contribute

chroma_datasets is currently backed by Hugging Face datasets

How to use

The following will:

1. Download the 2022 State of the Union with pre-computed chunks and embeddings
2. Import it into Chroma

Try it yourself in this Colab Notebook.

import chromadb
from chromadb.utils import embedding_functions
from chroma_datasets import StateOfTheUnion
from chroma_datasets.utils import import_into_chroma

chroma_client = chromadb.Client()
openai_ef = embedding_functions.OpenAIEmbeddingFunction(
    api_key="API_KEY",
    model_name="text-embedding-ada-002"
)
collection = import_into_chroma(chroma_client=chroma_client, dataset=StateOfTheUnion, embedding_function=openai_ef)
result = collection.query(query_texts=["The United States of America"], n_results=1)
print(result)

Adding a New Dataset

We welcome new datasets!

These datasets can be anything generally useful to developer education for processing and using embeddings.

Datasets should be exported from a Chroma collection. See examples/example_export.ipynb for an example of how to create a dataset on Hugging Face (the default path)

Create a new dataset from a Chroma Collection

(more examples of this in examples/example_export.ipynb)

Install dependencies

pip install datasets huggingface_hub chromadb

Login into Hugging Face

huggingface-cli login

Upload an existing collection to Hugging Face ** Hugging Face requires the data to have a "split name" - I suggest using a default of "data" **

import chromadb
from chroma_datasets.utils import export_collection_to_hf_dataset
client = chromadb.PersistentClient(path="./chroma_data")
dataset = export_collection_to_hf_dataset(
    client=client, 
    collection_name="paul_graham_essay", 
    license="MIT")
dataset.push_to_hub(
    repo_id="chromadb/paul_graham_essay", 
    split="data")

Create a Dataset Class and add it to chroma_datasets/__init__.py

• Set the string name of the embedding function you used to embed the data, this will make it possible for users to use the embeddings. Please also customize the helpful error message so if users pass in no embedding function or the wrong one, they get help.

class PaulGrahamEssay(Dataset):
    """
        http://www.paulgraham.com/worked.html
    """
    hf_data = None
    hf_dataset_name = "chromadb/pg_essay"
    embedding_function = "OpenAIEmbeddingFunction"
    embedding_function_instructions = ef_instruction_dict[embedding_function]

Utility API Documentation

Many of these methods are purely conveneient. This makes it easy to save and load Chroma Collections to disk. See ./examples/example_export.ipynb for example use.

from chromadb.utils import (
    export_collection_to_hf_dataset,
    export_collection_to_hf_dataset_to_disk,
    import_chroma_exported_hf_dataset_from_disk,
    import_chroma_exported_hf_dataset
)

# Exports a Chroma collection to an in-memory HuggingFace Dataset
def export_collection_to_hf_dataset(chroma_client, collection_name, license="MIT"):

# Exports a Chroma collection to a HF dataset and saves to the path
def export_collection_to_hf_dataset_to_disk(chroma_client, collection_name, path, license="MIT"):

# Imports a HuggingFace Dataset into a Chroma Collection
def import_chroma_exported_hf_dataset(chroma_client, dataset, collection_name, embedding_function=None):

# Imports a HuggingFace Dataset from Disk and loads it into a Chroma Collection
def import_chroma_exported_hf_dataset_from_disk(chroma_client, path, collection_name, embedding_function=None):

License

Code: Apache 2.0

Each dataset has it's own license. Datasets uploaded by Chroma are released as MIT.

Todo

• Add test suite to test some of the critical paths
• Add automated pypi release
• Add loaders for other locations (remote like S3, local like CSV... etc)
• Super easy streamlit/gradio wrapper to push up a collection to interact with