Common Datasets in Snapshot Format
You may find that creating embeddings from datasets is a very resource-intensive task. If you need a practice dataset, feel free to pick one of the ready-made snapshots on this page. These snapshots contain pre-computed vectors that you can easily import into your Qdrant instance.
Available Datasets
Our snapshots are usually generated from publicly available datasets, which are often used for non-commercial or academic purposes. The following datasets are currently available. Please click on a dataset name to see its detailed description.
| Dataset | Model | Vector size | Documents | Size | Qdrant snapshot | HF Hub |
|---|---|---|---|---|---|---|
| Qdrant Documentation | multilingual-e5-small | 384 | 3.6K | 24 MB | Download | Open |
| Amazon ESCI Products | multilingual-e5-small | 384 | 165K | 488 MB | Download | Open |
| Amazon Berkeley Objects | clip-ViT-B-32 | 512 | 45K | 183 MB | Download | Open |
| Arxiv.org Titles | InstructorXL | 768 | 2.3M | 7.7 GB | Download | Open |
| Arxiv.org Abstracts | InstructorXL | 768 | 2.3M | 9.2 GB | Download | Open |
| Wolt Food | clip-ViT-B-32 | 512 | 1.7M | 6.8 GB | Download | Open |
Once you download a snapshot, you need to restore it using the Qdrant CLI upon startup or through the API.
Qdrant on Hugging Face
Hugging Face provides a platform for sharing and using ML models and datasets. Qdrant is one of the organizations there! We aim to provide you with datasets containing neural embeddings that you can use to practice with Qdrant and build your applications based on semantic search. Please let us know if you’d like to see a specific dataset!
If you are not familiar with Hugging Face datasets, or would like to know how to combine it with Qdrant, please refer to the tutorial.
Qdrant Documentation
This dataset contains the Qdrant documentation, courses, and articles from qdrant.tech,
as of October 25, 2026, split into 3,593 chunks by page section. A section that fits in 480 tokens becomes
one chunk, and larger sections are split further. Each point stores a dense vector named dense and a
payload with the chunk text, its page title and URL, the section breadcrumb, the content type
(documentation, course, or article), and whether the chunk contains code:
{
"url": "/course/essentials/day-6/final-project/",
"title": "Final Project: Production-Ready Documentation Search Engine",
"section": "Build Steps > Step 3: Parse and Chunk Documents",
"content_type": "course",
"text": "Final Project: Production-Ready Documentation Search Engine > Build Steps > Step 3: Parse and Chunk Documents\n\nPick a documentation set and parse it into structured sections. Preserve the hierarchy users expect.\n\n...",
"has_code": false
}
The content_type, url, and has_code fields have payload indexes for filtering, and text has a full-text index with BM25 scoring for keyword and hybrid search. The index uses the code tokenizer on both the prose and the code, so identifiers match across naming styles: efConstruct finds ef_construct. Where a page shows the same code example in several client languages, only the Python version is kept.
The embeddings were generated with the multilingual-e5-small model from the text field with fenced
code blocks, Markdown images, and link URLs removed, using the passage: prefix. Embed your search
queries with the query: prefix instead, and pass using="dense" when you query. On Qdrant Cloud,
this model is available for free through Cloud Inference as
intfloat/multilingual-e5-small, so you can embed queries without hosting it yourself. The following
code snippet shows how to generate embeddings using the multilingual-e5-small model:
from sentence_transformers import SentenceTransformer
model = SentenceTransformer("intfloat/multilingual-e5-small")
text = "Final Project: Production-Ready Documentation Search Engine > Build Steps > Step 3: Parse and Chunk Documents"
embeddings = model.encode([f"passage: {text}"], normalize_embeddings=True)
query_embedding = model.encode(["how to chunk documentation"], prompt="query: ", normalize_embeddings=True)
Download the snapshot of the dataset. It requires Qdrant 1.20 or later.
Importing the Dataset
The easiest way to use the provided dataset is to recover it via the API by passing the URL as a location. It works also in Qdrant Cloud. The following code snippet shows how to create a new collection and fill it with the snapshot data:
PUT /collections/{collection_name}/snapshots/recover
{
"location": "https://snapshots.qdrant.io/qdrant-docs-2026-10-25.snapshot"
}
Amazon ESCI Products
This dataset contains 164,900 product listings from the Amazon Shopping Queries Dataset
(ESCI): every product judged for the US test queries of the dataset’s small version. Each point represents
a product and stores a dense vector named dense. The payload holds the product ID (ASIN), title, brand,
and color where available, along with the text used to create the embedding: the title, the brand, and
the product bullet points.
{
"product_id": "B01MT0QDPO",
"title": "Nature Valley Sweet and Salty Nut Granola Bars, Cashew, 12 ct, 14.8 oz",
"brand": "Nature Valley",
"text": "Nature Valley Sweet and Salty Nut Granola Bars, Cashew, 12 ct, 14.8 oz\nBrand: Nature Valley\nGRANOLA BAR: Nature Valley granola bars are made with whole grain oats roasted peanuts and peanut butter\n..."
}
The product_id, brand, and color fields have payload indexes for filtering, and text
has a full-text index with BM25 scoring for keyword and hybrid search. The snapshot contains the
products only. The 8,956 search queries and their 181,701 relevance labels are in the judgments
configuration of the Hugging Face dataset.
They are the rows of the source repository with
product_locale set to us, small_version set to 1, and split set to test.
The embeddings were generated with the multilingual-e5-small model from the text field with the
passage: prefix. Embed your search queries with the query: prefix instead, and pass using="dense"
when you query. On Qdrant Cloud, this model is available for free through
Cloud Inference as intfloat/multilingual-e5-small. The following code
snippet shows how to generate embeddings using the multilingual-e5-small model:
from sentence_transformers import SentenceTransformer
model = SentenceTransformer("intfloat/multilingual-e5-small")
text = "Nature Valley Sweet and Salty Nut Granola Bars, Cashew, 12 ct, 14.8 oz\nBrand: Nature Valley"
embeddings = model.encode([f"passage: {text}"], normalize_embeddings=True)
query_embedding = model.encode(["cashew granola bars"], prompt="query: ", normalize_embeddings=True)
Download the snapshot of the dataset. It requires Qdrant 1.20 or later. The Shopping Queries Dataset is provided by Amazon Science under the Apache 2.0 license.
Importing the Dataset
The easiest way to use the provided dataset is to recover it via the API by passing the URL as a location. It works also in Qdrant Cloud. The following code snippet shows how to create a new collection and fill it with the snapshot data:
PUT /collections/{collection_name}/snapshots/recover
{
"location": "https://snapshots.qdrant.io/esci-products.snapshot"
}
Amazon Berkeley Objects
This dataset contains 44,932 products from the Amazon Berkeley Objects (ABO) dataset. Each point represents a product with its main image, represented as a single unnamed vector of 512 float numbers. Products that share a main image, usually size or color variants, are kept once, and products whose main image is a placeholder rather than a photo are left out. The payload holds the item ID, product name, brand, color where available, product type, marketplace country, and the image URL:
{
"item_id": "B0716KHKRX",
"name": "Amazon Brand - Solimo Travel Case for Small Electronics and Accessories (Red)",
"brand": "Amazon Brand - Solimo",
"color": "Red",
"product_type": "PORTABLE_ELECTRONIC_DEVICE_COVER",
"country": "IN",
"image_url": "https://amazon-berkeley-objects.s3.amazonaws.com/images/small/f3/f3de0cfd.jpg"
}
The item_id, brand, color, product_type, and country fields have payload indexes for
filtering, and name has a full-text index with BM25 scoring.
The embeddings were generated with the clip-ViT-B-32 model from the image at image_url, using the
following code snippet. Because CLIP maps text and images into the same space, you can search the images
with text queries encoded by the same model. On Qdrant Cloud, the same models are available through
Cloud Inference as qdrant/clip-vit-b-32-vision and qdrant/clip-vit-b-32-text:
from PIL import Image
from sentence_transformers import SentenceTransformer
image_path = "f3de0cfd.jpg"
model = SentenceTransformer("clip-ViT-B-32")
embedding = model.encode(Image.open(image_path))
query_embedding = model.encode("a red travel case")
Download the snapshot of the dataset. It requires Qdrant 1.20 or later. ABO is provided by Amazon.com and UC Berkeley under the CC BY 4.0 license. This collection is a modified subset: products with an English name and a photo as their main image, one product per main image, and 1,000 of the phone cases, with a CLIP embedding of each main image added.
Importing the Dataset
The easiest way to use the provided dataset is to recover it via the API by passing the URL as a location. It works also in Qdrant Cloud. The following code snippet shows how to create a new collection and fill it with the snapshot data:
PUT /collections/{collection_name}/snapshots/recover
{
"location": "https://snapshots.qdrant.io/amazon-berkeley-objects.snapshot"
}
Arxiv.org
Arxiv.org is a highly-regarded open-access repository of electronic preprints in multiple fields. Operated by Cornell University, arXiv allows researchers to share their findings with the scientific community and receive feedback before they undergo peer review for formal publication. Its archives host millions of scholarly articles, making it an invaluable resource for those looking to explore the cutting edge of scientific research. With a high frequency of daily submissions from scientists around the world, arXiv forms a comprehensive, evolving dataset that is ripe for mining, analysis, and the development of future innovations.
Arxiv.org Titles
This dataset contains embeddings generated from the paper titles only. Each vector has a payload with the title used to create it, along with the DOI (Digital Object Identifier).
{
"title": "Nash Social Welfare for Indivisible Items under Separable, Piecewise-Linear Concave Utilities",
"DOI": "1612.05191"
}
The embeddings generated with InstructorXL model have been generated using the following instruction:
Represent the Research Paper title for retrieval; Input:
The following code snippet shows how to generate embeddings using the InstructorXL model:
from InstructorEmbedding import INSTRUCTOR
model = INSTRUCTOR("hkunlp/instructor-xl")
sentence = "3D ActionSLAM: wearable person tracking in multi-floor environments"
instruction = "Represent the Research Paper title for retrieval; Input:"
embeddings = model.encode([[instruction, sentence]])
Download the snapshot of the dataset.
Importing the Dataset
The easiest way to use the provided dataset is to recover it via the API by passing the URL as a location. It works also in Qdrant Cloud. The following code snippet shows how to create a new collection and fill it with the snapshot data:
PUT /collections/{collection_name}/snapshots/recover
{
"location": "https://snapshots.qdrant.io/arxiv_titles-2108082541245612-2026-06-04-09-55-52.snapshot"
}
Arxiv.org Abstracts
This dataset contains embeddings generated from the paper abstracts. Each vector has a payload with the abstract used to create it, along with the DOI (Digital Object Identifier).
{
"abstract": "Recently Cole and Gkatzelis gave the first constant factor approximation\nalgorithm for the problem of allocating indivisible items to agents, under\nadditive valuations, so as to maximize the Nash Social Welfare. We give\nconstant factor algorithms for a substantial generalization of their problem --\nto the case of separable, piecewise-linear concave utility functions. We give\ntwo such algorithms, the first using market equilibria and the second using the\ntheory of stable polynomials.\n In AGT, there is a paucity of methods for the design of mechanisms for the\nallocation of indivisible goods and the result of Cole and Gkatzelis seemed to\nbe taking a major step towards filling this gap. Our result can be seen as\nanother step in this direction.\n",
"DOI": "1612.05191"
}
The embeddings generated with InstructorXL model have been generated using the following instruction:
Represent the Research Paper abstract for retrieval; Input:
The following code snippet shows how to generate embeddings using the InstructorXL model:
from InstructorEmbedding import INSTRUCTOR
model = INSTRUCTOR("hkunlp/instructor-xl")
sentence = "The dominant sequence transduction models are based on complex recurrent or convolutional neural networks in an encoder-decoder configuration. The best performing models also connect the encoder and decoder through an attention mechanism. We propose a new simple network architecture, the Transformer, based solely on attention mechanisms, dispensing with recurrence and convolutions entirely. Experiments on two machine translation tasks show these models to be superior in quality while being more parallelizable and requiring significantly less time to train."
instruction = "Represent the Research Paper abstract for retrieval; Input:"
embeddings = model.encode([[instruction, sentence]])
Download the snapshot of the dataset.
Importing the Dataset
The easiest way to use the provided dataset is to recover it via the API by passing the URL as a location. It works also in Qdrant Cloud. The following code snippet shows how to create a new collection and fill it with the snapshot data:
PUT /collections/{collection_name}/snapshots/recover
{
"location": "https://snapshots.qdrant.io/arxiv_abstracts-2108082541245612-2026-06-04-09-56-06.snapshot"
}
Wolt Food
Our Food Discovery demo relies on the dataset of food images from the Wolt app. Each point in the collection represents a dish with a single image. The image is represented as a vector of 512 float numbers. There is also a JSON payload attached to each point, which looks similar to this:
{
"cafe": {
"address": "VGX7+6R2 Vecchia Napoli, Valletta",
"categories": ["italian", "pasta", "pizza", "burgers", "mediterranean"],
"location": {"lat": 35.8980154, "lon": 14.5145106},
"menu_id": "610936a4ee8ea7a56f4a372a",
"name": "Vecchia Napoli Is-Suq Tal-Belt",
"rating": 9,
"slug": "vecchia-napoli-skyparks-suq-tal-belt"
},
"description": "Tomato sauce, mozzarella fior di latte, crispy guanciale, Pecorino Romano cheese and a hint of chilli",
"image": "https://wolt-menu-images-cdn.wolt.com/menu-images/610936a4ee8ea7a56f4a372a/005dfeb2-e734-11ec-b667-ced7a78a5abd_l_amatriciana_pizza_joel_gueller1.jpeg",
"name": "L'Amatriciana"
}
The embeddings generated with clip-ViT-B-32 model have been generated using the following code snippet:
from PIL import Image
from sentence_transformers import SentenceTransformer
image_path = "5dbfd216-5cce-11eb-8122-de94874ad1c8_ns_takeaway_seelachs_ei_baguette.jpeg"
model = SentenceTransformer("clip-ViT-B-32")
embedding = model.encode(Image.open(image_path))
Download the snapshot of the dataset.
Importing the Dataset
The easiest way to use the provided dataset is to recover it via the API by passing the URL as a location. It works also in Qdrant Cloud. The following code snippet shows how to create a new collection and fill it with the snapshot data:
PUT /collections/{collection_name}/snapshots/recover
{
"location": "https://snapshots.qdrant.io/wolt-clip-2108082541245612-2026-06-04-09-56-17.snapshot"
}