SlideOn

SlideOn

An English sparse retrieval model from LightOn that turns queries and documents into sparse vectors for inverted-index search. It suits search engineers and developers building retrieval over English text.

Made by lightonai on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: SlideOn is an English sparse retrieval model from LightOn with 174 million parameters. It turns each query and each document into a sparse vector with 32,768 dimensions, so a search system can keep the vectors in an inverted index and rank matches by dot product. It starts from LightOn's DenseOn model and was trained with sparse autoencoder pretraining, then retrieval fine-tuning.

Who it suits: It suits search engineers and developers who retrieve English text and want vectors that fit a classic inverted index. Researchers who compare sparse and dense retrieval can use it as one more data point.

Title: Who might use it
Industry | Who uses it | How they use it | Example
Search | Search engineers | Store sparse vectors in an inverted index | Keep each document's active latents and look up posting lists for a query
AI applications | Developers building retrieval for language models | Fetch English passages for a question | Pull the best matching documents before an answer is written
Research | Retrieval researchers | Compare sparse and dense retrieval | Test the model on the BEIR-15 evaluation sets

If your text is multilingual, long, or mostly code, you can skip it. The card points to its sibling mSlideOn for those cases.

Where to use it: The record lists sentence-transformers as its library, so Python is the natural place. The sparse vectors also suit any search system built on an inverted index.

• Python with the sentence-transformers library
• Search systems that use an inverted index
• Windows, macOS or Linux with Python, to download the files

Title: Jobs it does
Use case | Who | What they get back
English document search | Search engineers | Ranked documents scored by dot product
Cheaper storage and search | Platform teams | Vectors you can prune at query time without retraining
Retrieval research | Researchers | Scores to compare against dense models

How to set it up: You need Python with pip and the sentence-transformers library. Fetch the files from Hugging Face, then follow the model card for the exact loading calls.

• Python with pip
• The sentence-transformers library
• Enough disk space for roughly 697 MB of weights in F32

1. Install the Hugging Face download tool.
2. Download the model files to your computer.
3. Load the model through sentence-transformers, following its card.

pip install -U huggingface_hub
hf download lightonai/SlideOn

How to use it: Encode your queries and documents with the model to get sparse vectors. Store the active dimensions of each document in an inverted index. At search time, the query's dimensions select posting lists and their weights add up into document scores. Inputs can run to 8,192 tokens, and query and document sparsity can be set separately.

Example: A developer indexes a set of English documents, then encodes a user question with the model. A dot-product search over the index returns the best matching documents. With queries capped at 64 nonzero dimensions and documents at 512, the card's score stays almost the same.

For newcomers and experts: If you are new, sparse retrieval works like keyword search, except the model learns the keywords instead of reading them off the page. Start by downloading the files and reading the card's section on encoding.

If you know the field, a learned linear encoder maps each DenseOn token representation to its top 32 activations, followed by ReLU. Max pooling across tokens gives one vector per text. Joint query and document Matryoshka training keeps quality under pruning, so you tune cost at inference time.

Query or document text comes in
DenseOn token representations are built
A learned encoder keeps the top 32 activations
Max pooling gives one sparse vector
A dot product over the index ranks documents

Good to know: The card reports BEIR-15 scores as average nDCG@10 over 15 datasets, higher is better. Capping nonzero values trims retrieval cost with little change in score.

Title: BEIR-15 score, nDCG@10
SlideOn unpruned: 56.68
SlideOn docs capped at 512: 56.98
SlideOn queries 64, docs 512: 56.66
DenseOn: 56.20

• License: Apache 2.0
• Cost to run: none named
• Data it reaches: the model files from Hugging Face, and the text you choose to encode
• Not for: multilingual, long-context or code retrieval, where the card points to mSlideOn

On Hugging Face, 13 November 2026

• Task: feature extraction
• Library: sentence-transformers
• Parameters: 174M
• Weights: about 697 MB in F32; about 87 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 830
• Likes: 18
• License: apache-2.0
• Languages: en
• Based on: lightonai/DenseOn
• Quantized versions: 1
• Tags: sentence-transformers, safetensors, slideon, sparse, sentence-similarity, custom-code, sparse-retrieval, matryoshka, feature-extraction

Links

• Model: https://huggingface.co/lightonai/SlideOn
• Files: https://huggingface.co/lightonai/SlideOn/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download lightonai/SlideOn

Then follow the usage in its model card. Nothing in this kit runs by itself.