ML Frameworks

VisionHOPE is a PyTorch visual backbone built on self-referential nested learning, where what the model remembers and how it learns co-evolve while it processes a single image.

It replaces autoregressive LLM classification with a single parallel forward pass, cutting latency from seconds to sub-20 ms while staying local.

Most 'pick one of these' LLM calls waste a generative model on a classifier's job — Jeff fine-tunes small open models to return a calibrated probability per option in a single forward pass, on your own hardware.

NeoHorse-1 prototypes a feedback loop where a routing harness evaluates agent trajectories and feeds capability-level signals back into the training mixture, aiming for recursive self-improvement.

Halo scales LLM and multimodal training from one GPU to multi-node clusters by wrapping standard Hugging Face models in-place instead of maintaining a separate distributed rewrite.

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

DanKS is Kingsoft’s attempt to master a four-player Chinese card game using structured retrieval and PPO self-play, shipped as a complete research lineage.

Spirula Studio rebuilds the entire 3D Gaussian Splatting pipeline—video to splat to textured mesh—as a single C++ binary that runs on NVIDIA, AMD, Intel, and Apple GPUs via Vulkan, no Python required.

TabICL treats training and prediction as a single transformer forward pass, aiming to eliminate hyperparameter tuning while matching or beating heavily tuned gradient-boosted baselines on most TabArena datasets.

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.

DiffusionOPSD closes the supervision gap in diffusion reward optimization by distilling explicit, continually refreshed targets from the model's own behavior policy.

This is the PyTorch library and official training scripts AllenAI uses to build its open OLMo language models.

OmniChar Studio packages a character's face, body, and wardrobe into a portable `.char` file that stays consistent across models — and trains the LoRAs to back it up.

mjbatch exists so you can run thousands of MuJoCo simulations in parallel on CPU without Python’s GIL getting in the way.

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

KubeRay turns distributed Python workloads into ordinary Kubernetes resources, with lifecycle, autoscaling, and zero-downtime serving handled for you.

OneTrainer wraps fine-tuning, LoRA, and dataset prep for a huge pile of diffusion models into one desktop app — no notebook archaeology required.

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.

Fizgig trains LoRAs on 8 GB cards, but its real pitch is the workbench: repair, remix, and profile LoRAs you already have.

Google ships two skills that teach AI coding agents how to actually work with Gemma models.