Projects sharing features with Olmocr

This project is a self-hosted email verification system and API designed to validate email existence and clean mailing lists on private infrastructure. It functions as a deliverability tool that confirms if email addresses are reachable by communicating with mail servers via the SMTP protocol without sending actual messages. The system is distinguished by its high-volume SMTP infrastructure, which utilizes a stateless worker architecture and message queue task distribution to scale validation tasks. It includes an SMTP proxy gateway that routes requests through SOCKS5 proxies to mask server i

This project is an AI-powered document processing engine designed to transform diverse file formats into structured Markdown. By leveraging multimodal language models, it performs complex layout analysis and semantic text extraction, allowing for the conversion of both unstructured files and scanned images into machine-readable content. The toolkit distinguishes itself through a modular, plugin-based architecture that orchestrates multi-stage extraction pipelines. Users can steer the parsing behavior by injecting custom instructions, enabling the system to adapt to domain-specific document st

Easy-dataset is a comprehensive platform designed for the end-to-end management of machine learning datasets, specifically tailored for language and vision model fine-tuning. It functions as a centralized environment for the entire data lifecycle, encompassing the automated generation of synthetic training data, the structural organization of document collections, and the systematic annotation of individual data points. The platform distinguishes itself through its integrated evaluation and orchestration capabilities. It provides a dedicated suite for benchmarking models, featuring blind side

This project provides a collection of Helm charts designed for deploying and managing self-hosted error tracking and performance monitoring infrastructure within a Kubernetes cluster. It enables the deployment of complex, multi-service application stacks using standardized configuration templates to ensure consistent and repeatable infrastructure management. The platform distinguishes itself by automating the operational requirements of distributed event processing, including the horizontal autoscaling of background task workers and the secure generation of cryptographic keys for internal ser

Light Task Scheduler is a distributed job scheduling and workflow orchestration platform designed for managing background processing across scalable computing environments. It functions as a cluster management system that coordinates stateless nodes to execute recurring, cron-based, or one-time tasks with centralized control and high availability. The platform distinguishes itself through a leader-based coordination model that automatically elects a primary controller to manage task distribution and system state. It supports complex workflow dependencies, ensuring that prerequisite tasks comp

This project provides a containerized environment for deploying Apache Airflow, enabling the orchestration of complex data pipelines and automated task scheduling. By packaging the orchestration platform into portable images, it ensures consistent execution across diverse infrastructure setups and simplifies the management of runtime dependencies. The platform facilitates distributed task execution by decoupling the scheduler from the execution layer, allowing for horizontal scaling of processing power across multiple worker nodes. It supports dynamic configuration through environment variabl

FastChat is a training and serving platform for large language models that provides an integrated toolkit for fine-tuning, hosting, and benchmarking chatbots. It functions as an inference server capable of hosting multiple models and exposing them via a standardized API for chat applications. The platform distinguishes itself through a distributed model controller that manages worker nodes and routes requests across a hardware-agnostic inference layer supporting various accelerators. It includes a dedicated evaluation framework for assessing model quality using automated judges, multi-turn di

NATS Server is a high-performance, lightweight messaging system designed for cloud-native applications, edge computing, and distributed microservices. It functions as a distributed publish-subscribe broker that routes messages using hierarchical, dot-separated subject strings, enabling decoupled communication between services without requiring centralized broker lookups. The system supports core messaging patterns including asynchronous publish-subscribe, request-reply, and load-balanced queue processing. The platform distinguishes itself through a decentralized architecture that eliminates t

LLaMA-Factory is a comprehensive suite for dataset preparation, model fine-tuning, memory optimization, and standardized API deployment. It provides a unified platform for the supervised and reward-based fine-tuning of large language models and vision-language models. The framework includes a specialized toolkit for training vision-language models and a model serving interface that deploys trained models through high-performance APIs. It utilizes precision tuning and quantization techniques to reduce the hardware requirements and memory footprint of large models. The system covers data pipel

h2oGPT is a self-hosted platform designed for running large language models and executing retrieval-augmented generation workflows locally. It provides a comprehensive web interface that allows users to index private document collections into searchable databases, enabling context-aware question answering and summarization without exposing sensitive data to external services. The platform distinguishes itself by offering a modular architecture that supports both local model execution and connections to external inference servers. It facilitates the development of autonomous agents capable of

This project is a comprehensive toolkit designed for the full lifecycle management of large language and multimodal models. It functions as a unified orchestrator that handles the entire development process, ranging from dataset preparation and supervised fine-tuning to advanced reinforcement learning alignment and production-ready inference deployment. The platform distinguishes itself through a specialized reinforcement learning library that supports complex optimization algorithms, including group relative policy optimization and leave-one-out techniques, to improve model instruction-follo

Jo is a command-line utility designed to construct and manipulate JSON objects and arrays directly from shell arguments and standard input. It functions as a data processing tool that transforms raw input into structured formats, enabling the generation of complex payloads for APIs, configuration files, and automated data pipelines. The tool distinguishes itself through its ability to resolve hierarchical data structures using delimiter-based path definitions and its integrated type-inference engine, which automatically casts input values into native boolean, numeric, or null types. Users can

OCRFlux is a lightweight yet powerful multimodal toolkit that significantly advances PDF-to-Markdown conversion, excelling in complex layout handling, complicated table parsing and cross-page content merging.

BabelDOC is a technical document translation system designed to translate PDF files while preserving their original layout and styling. It functions as a layout-preserving translator that utilizes large language models to convert content into target languages, specifically tailored for scientific and technical documents. The system distinguishes itself through specialized handling of academic content, including the identification and preservation of mathematical formulas and complex layout structures. It ensures technical accuracy by employing glossary-driven terminology enforcement, using so

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

Sparrow is an LLM document extraction platform and vision-based inference engine designed to convert images and PDFs into validated structured data. It functions as an agentic workflow orchestrator that chains classification, extraction, and validation tasks into multi-step pipelines. The system distinguishes itself through a backend-agnostic inference layer that manages models across local GPUs, Apple Silicon, and cloud providers. It employs coordinate-based visual grounding to map extracted text to precise bounding box coordinates and utilizes hint-based model steering to guide attention an

MinHash, LSH, LSH Forest, Weighted MinHash, HyperLogLog, HyperLogLog++, LSH Ensemble and HNSW

Manage scalable open LLM inference endpoints in Slurm clusters

This tool is a command-line processor designed for querying, updating, and transforming structured data files. It functions as a versatile engine for manipulating YAML, JSON, TOML, and XML documents, allowing users to perform complex operations directly from the terminal. By utilizing a path-based expression language, it enables precise navigation and modification of data structures within configuration files and infrastructure-as-code workflows. What distinguishes this tool is its ability to perform in-place document mutations while preserving original formatting, comments, and metadata. It

Code for the paper "Rethinking Benchmark and Contamination for Language Models with Rephrased Samples"

omniparser: a native Golang ETL streaming parser and transform library for CSV, JSON, XML, EDI, text, etc.

sChandra is a document processing platform that converts images, PDFs, Word documents, spreadsheets, and other formats into structured output such as HTML, Markdown, or JSON while preserving layout. It can also extract specific data fields from invoices, contracts, or reports using user-defined JSON schemas, with citations back to source locations. The service supports form filling in PDF and image documents, document generation from Markdown, and extraction of tracked changes from Word files. The platform distinguishes itself with pipeline-based processing chains that combine multiple proces

Fast Multimodal Semantic Deduplication & Filtering

Docling is a multimodal content converter and document parser designed to transform PDFs, Office files, and HTML into structured Markdown or JSON for generative AI applications. It functions as an OCR document processor and a PDF layout analyzer that extracts tables, charts, and hierarchical structures while preserving the original page layout. The system operates as a local-first inference engine, allowing for the processing of sensitive data in air-gapped environments without external network connectivity. It can also be deployed as an API or a Model Context Protocol server to provide parsi