#python
Verifiers is a tool that helps create environments for training large language models (LLMs) using reinforcement learning (RL). It includes features like async GRPO training and integration with other frameworks. This tool is useful for building and evaluating LLMs in various tasks, such as creating synthetic data or using tools within models. It supports both single-turn and multi-turn interactions, making it versatile for different applications. By using Verifiers, users can efficiently train and evaluate LLMs, which can improve their performance and accuracy in tasks like answering questions or generating text.
https://github.com/willccbb/verifiers
#python
Related Website Sets (RWS) let companies declare which of their websites are related, so browsers like Chrome can allow limited sharing of cookies across these sites. This helps keep features like staying signed in or remembering your shopping cart working smoothly when you move between related sites, even as browsers block most third-party cookies to protect your privacy. By submitting their sites as a set, companies ensure a better, more seamless experience for you without losing privacy protections. This system balances user convenience with privacy by controlling when cross-site data is shared.
https://github.com/GoogleChrome/related-website-sets
#python
Nemo RL is a powerful and scalable library that helps you efficiently train and fine-tune large AI models, from small ones on a single GPU to huge models with over 100 billion parameters using thousands of GPUs. It integrates easily with Hugging Face models and uses NVIDIA’s Megatron Core for fast, optimized training with advanced parallelism, making it ideal for very large models and long sequences. It supports various reinforcement learning algorithms and fine-tuning methods, offers flexible resource management with Ray, and provides detailed, user-friendly documentation and examples. This means you can train state-of-the-art AI models faster and more reliably, even at massive scale, with less hassle.
https://github.com/NVIDIA-NeMo/RL
#python
Terminal-Bench is a tool that tests how well AI agents perform real tasks in a terminal, like compiling code or setting up servers, all on their own. It includes a set of tasks with instructions and tests, plus a system that connects AI models to a safe terminal environment. You can install it easily with pip and run tests to see how good your AI is at practical, real-world jobs. This helps you build, compare, and improve AI agents for coding and system tasks, making your AI development more reliable and measurable. You can also contribute new tasks or join a leaderboard to track progress.
https://github.com/laude-institute/terminal-bench
#python
AI Toolkit by Ostris is a powerful, easy-to-use software suite for training and fine-tuning AI models like Stable Diffusion and FLUX.1 on consumer-grade Nvidia GPUs with at least 24GB VRAM. It supports image and video models, offers both graphical and command-line interfaces, and allows training specific neural network layers. You can run it locally or on cloud platforms like RunPod and Modal. The toolkit automatically handles dataset preparation and resizing, and includes a web UI for managing training jobs securely. This helps you efficiently create custom AI models with flexibility and advanced features, even if you are not an expert.
https://github.com/ostris/ai-toolkit
#python
Buttercup is an AI-powered system that automatically finds and fixes security bugs in open-source software by running smart tests (fuzzing) and then creating patches to fix vulnerabilities. It works mainly on Linux with at least 8 CPU cores, 16 GB RAM, and 100 GB disk space, using AI models from providers like OpenAI or Anthropic. You can easily set it up on your computer, monitor its progress through a web interface, and control costs by limiting AI usage. This helps you improve software security efficiently without manual code review, saving time and reducing risks from software bugs.
https://github.com/trailofbits/buttercup
#python
SkyReels-V2 is an advanced open-source video generation model that can create realistic, long-length, and film-style videos without a fixed time limit. It uses a new "Diffusion Forcing" method that lets it generate videos of theoretically infinite length, supporting both text-to-video and image-to-video creation. The model combines large language models, multi-stage training, and reinforcement learning to improve video quality, motion realism, and prompt accuracy. It also includes SkyCaptioner-V1, a powerful video captioning tool for better video understanding. This means you can generate high-quality, coherent videos for storytelling, animation, or film projects on your own computer or cloud, with flexible control over video length and content.
https://github.com/SkyworkAI/SkyReels-V2
#python
You can use ComfyUI wrapper nodes to easily try out WanVideo and related AI video models without dealing with complex core code. This setup lets you quickly test new models and features in a flexible, experimental way, even if they are not yet available in ComfyUI’s native system. It supports installing various WanVideo models and extras, including advanced video generation and editing tools, with simple steps for setup. This helps you explore creative video generation options, test new releases, and use powerful models like Wan2.1 for text-to-video or image-to-video tasks, all while managing resources efficiently on consumer GPUs.
https://github.com/kijai/ComfyUI-WanVideoWrapper
#python
TransformerLens is a Python library that helps you understand how GPT-2 style language models work inside by showing you their internal processes and activations. You can load over 50 open-source models, run them on text, and see or change what happens inside the model step-by-step. This makes it easier to study and reverse engineer the model’s learned algorithms without needing huge computing power. It’s great for researchers or anyone curious about how language models think, offering tutorials and tools to explore model behavior in detail. You just install it with pip and start analyzing models quickly. This helps you learn, debug, or improve language models effectively.
https://github.com/TransformerLensOrg/TransformerLens
#python
The Temporal Python SDK lets you write reliable, scalable workflows and activities in Python for long-running business processes. It supports type safety, async and threaded activities, and integrates with Python’s asyncio for smooth task management. You can run workflows that handle retries, failures, and cancellations automatically, making your code fault-tolerant without extra effort. The SDK includes tools for testing, debugging, and monitoring, plus a sandbox to avoid non-deterministic code. It also now supports integration with OpenAI Agents for building AI workflows. This helps you build complex, resilient applications faster and with less code.
https://github.com/temporalio/sdk-python
#python
LocalGPT lets you chat with your own documents securely on your computer without sending any data online, keeping your information completely private. It supports many open-source AI models and works on different devices like GPUs, CPUs, and Apple silicon. You can upload various file types, ask questions, and get answers locally, even without internet after initial setup. It also remembers your chat history during sessions and offers a user-friendly interface and API for developers. This means you get powerful AI help while fully controlling your data and avoiding privacy risks or subscription fees. It’s ideal for anyone needing secure, offline AI document interaction[2][1][4].
https://github.com/PromtEngineer/localGPT
#python
Agentic Document Extraction is a Python library that makes it easy to pull out structured data—like tables, charts, and text—from complex documents such as long PDFs or images, using advanced AI that understands both the content and layout. It automatically splits big files, processes them in parallel, and handles errors or rate limits, so you don’t have to worry about technical details. The benefit is that you get fast, accurate, and organized results—ready for analysis or automation—without needing to write lots of code or manage tricky document formats yourself[1][2][3].
https://github.com/landing-ai/agentic-doc