LightOn ORG

Late Interaction Models Training & Retrieval

NextPlaid, ColGREP: Multi-vector search, from database to coding agents.

High-Performance Engine for Multi-Vector Search

RITA is a family of autoregressive protein models, developed by LightOn in collaboration with the OATML group at Oxford and the Debora Marks Lab at Harvard.

Study on the applicability of Direct Feedback Alignment to neural view synthesis, recommender systems, geometric learning, and natural language processing.

PyLate efficient inference engine

Efficient BM25 with DuckDB 🦆

A fast, streaming-friendly BM25 search engine in Rust with mmap support

Implementation of NEWMA: a new method for scalable model-free online change-point detection

Python library for running large-scale computations on LightOn's OPUs

Double Descent Curve with Optical Random Features

Optical Transfer Learning

Code for our paper on best practices to train neural networks with direct feedback alignment (DFA).

ML benchmarks performance featuring LightOn's Optical Processing Unit (OPU) vs CPU and GPU.

Inference code in Pytorch for GPT-like models, such as PAGnol, a family of models with up to 1.5B parameters, trained on datasets in French.

Python implementation of supervised PCA, supervised random projections, and their kernel counterparts.

A Rust rewrite of FastKMeans for CPU-based clustering

Code to perform Model-Free Episodic Control using Aurora OPUs

Conformational exploration SARS-CoV-2 (coronavirus responsible for COVID-19)

Architecture embeddings independent from the parametrization of the search space

Python client for the LightOn Muse API

Public rankings of extreme-scale models

Double Trouble in the Double Descent Curve with Optical Processing Units.

A new defense mechanism against adversarial attacks through Optical Processing Units and synthetic gradients.

Minimal code to perform phase retrieval on LightOn OPUs

Port of Mamba to run and run efficiently on AMD.

Supercharge Your Model Training

Distilabel is a framework for synthetic data and AI feedback for engineers who need fast, reliable and scalable pipelines based on verified research papers.

BrowseComp-Plus: A More Fair and Transparent Evaluation Benchmark of Deep-Research Agent

Port of causal-conv1d to run and run efficiently on AMD.

Freeing data processing from scripting madness by providing a set of platform-agnostic customizable pipeline processing blocks.

Homebrew tap for colgrep — semantic code search

Just here to get around some import issues with transformers. We need particular versions of transformers and it isn't compatible with the published package.

Source code for the integration of Muse in Google Sheets.

Documentation for the Muse API.

No description.

ParseBench - A Document Parsing Benchmark for AI Agents

Late interaction (multi-vector) retrieval for PyTerrier, backed by PyLate

Homebrew tap for LightOn tools

Demo LightOn API use case of a procurement document verification (DC4)

A high-throughput and memory-efficient inference and serving engine for LLMs

No description.

Multi-Turn RAG Benchmark

No description.

Speakeasy generated python SDK for Paradigm

TRL forked for RLVR

No description.

Structured Text Generation

Large Language Model Text Generation Inference

A blazing fast inference solution for text embeddings models

the AI-native open-source embedding database

Collection of examples for the Muse models.