ML Frameworks

SolarWM exists to make long-horizon video world models reproducible, open-sourcing the full stack from a unified data contract to staged training recipes and checkpoints across three major backbones.

Official code for a book that ships PicoAgents, a full multi-agent framework written from scratch so you can see how the pieces fit together without vendor abstraction getting in the way.

It consolidates Fetch, Shadow Hand, D4RL, and other MuJoCo robotics benchmarks into a single Gymnasium-compatible suite, replacing legacy mujoco-py with modern bindings.

A WebGPU library that treats shader files like typed TypeScript modules and runs identically in the browser, headless Node, or a deterministic mock for tests.

It exists because Persian TTS and linguistics tools need a modern, learned grapheme-to-phoneme converter that outputs standard IPA without brittle hand-written rules.

It rounds up pre-built Windows binaries for AI libraries that typically force users into complicated, error-prone source builds.

Omega-AI was built from scratch in Java so JVM-native developers can train neural nets, run YOLO, and even generate images without bridging into Python ecosystems.

Soup exists because fine-tuning LLMs shouldn't require a cloud budget, SSH, or a PhD in distributed systems.

A production-hardened fork of slime that keeps massive MoE models from collapsing by obsessing over bit-wise alignment between rollout and training.

A curated collection of Jupyter notebooks and R Markdown files covering the standard ML curriculum, useful mostly as a reference for how to structure your own.

It trains multi-step LLM agents by treating every turn as a step-level MDP transition instead of a single growing prompt-response sequence.

Molt exists because most RL frameworks are too large to hack or bury you in backend ceremony, so it strips agentic RL down to ~9.2K lines of PyTorch with a single-actor trainer and Python-native rewards.

It exists to give individual quants an integrated, locally hosted alternative to cloud platforms by packing Qlib, LLM-driven factor mining, and broker-grade execution into a single Docker stack.

To prove that a 2.78-trillion-parameter model can run on a single CPU with 8 GB of RAM and no GPU.

RLinf exists because fine-tuning policies for robots and agents still requires rewriting your training stack for every new simulator, world model, or hardware rig.

It split off from `verl` to give diffusion, video, and omni-modality models an RL post-training framework that doesn't treat them like chatbots.

PySR hunts for human-readable equations that fit your data, using a Julia-backed genetic search engine wrapped in a scikit-learn interface.

Curated technical deep-dives covering everything from NVLink signal integrity to Kubernetes GPU scheduling and Huawei NPU porting.

Chinese open-source community Datawhale built a from-zero embodied-AI course that gets you from `print('hello')` to fine-tuning SmolVLA and Pi0.

It turns LLM theory into a stack of runnable notebook exercises—problem sets, answers, and hardware benchmarks included.