Models

Three locally fine-tuned Russian-language models for a self-hosted RAG stack: a document splitter, a search-query expander and a retrieval embedder, distilled from a teacher model and quantized to GGUF for llama.cpp on AMD hardware.

• RAG
• Models

Research code that reproduces the experiments behind the paper 'MLPs are Hebbians: Constructing Efficient Fact-Storing MLPs for Transformers', including Hebbian MLP construction, Transformer integration and fact-editing experiments.

• Models
• Papers

A fine-tuning SDK that captures an AI agent's real traffic, trains a small open model to match it, and gradually shifts the agent from a costly frontier model to the model you now own.

• Models

Fine-tunes open-source language models using Tinker from inside the Pi agent, preparing training data, running small jobs and comparing the trained checkpoint against the original model.

• Models

A structured, markdown-based synthetic dataset covering business, technology, psychology and other topics, created to give AI training, testing and question-answering research a consistent, well-organized knowledge source.

• Models

A quantization-aware training framework built on Keras 3 that optimizes neural-network bit-widths by gradient, down to a per-weight and per-activation level, for real-time deep learning on FPGAs and other edge devices.

• Models

A governed, local-first system for training small language models that measures each new model against a benchmark, promotes only improvements through a regression-aware gate, and archives the evidence for every decision.

• Models
• AI evals

A 12-million-parameter transformer language model implemented in Node.js and trained with the mni-ml/framework library on BPE-tokenized text such as TinyStories or YouTube-Commons transcripts.

• Models

A deep learning tensor library written from scratch in pure C, with autograd, GPT-style transformer building blocks, and CPU-only SIMD-accelerated training with no external ML dependencies.

• Models

A platform that turns an AI agent's own production traces into a fine-tuned model, benchmarked against user-defined evaluation metrics and served behind one OpenAI-compatible API, available hosted or self-hosted.

• Models
• AI evals

A PyTorch library that trains a linear adapter with triplet loss to align retrieval query embeddings with real user queries, improving RAG search without retraining or re-indexing the embedding model.

• RAG
• Models

A gradient-free method for training neural networks, including spiking networks and other models without useful gradients, that perturbs parameters within a finite radius and updates them through a weighted optimal-transport barycenter.

• Models

A Python library and CLI that turns source documents such as PDFs into instruction fine-tuning datasets, grounding each generated example in a specific chunk of the source, grading it with an LLM judge, and deduplicating by embedding rather than exact text.

• Models

A cross-repository control project integrating Newton-Muon and Locoprop-style preconditioned optimizers, with FSDP and MuP support, into Megatron-LM and TransformerEngine for training mixture-of-experts models at scale.

• Models

A reference implementation of an optimizer for budget-constrained discrete model compression, which relaxes a per-layer choice of compression option onto a Riemannian manifold so a fixed overall budget, such as total bits, is satisfied exactly rather than only approximately.

• Models

A research codebase for fine-tuning, merging and transporting task vectors between vision and language models, including rebasin methods that align independently trained networks before merging them.

• Models

A 27B-parameter open-weight language model, built on Qwen3.8-27B, tuned specifically for everyday writing such as messages and emails, released free for non-commercial use under a CC BY-NC license.

• Models

A research technique that attaches a small hypernetwork to a recursive transformer block, generating a different LoRA modulation at each recurrence step instead of repeating the same transformation, improving depth-recycled models.

• Models

An end-to-end LLM toolkit that runs entirely on one consumer NVIDIA GPU, taking a user from environment diagnosis through pretraining, fine-tuning, evaluation, GGUF export and a self-contained benchmark report.

• Models
• Tutorials

An automated pipeline that converts Hugging Face models to GGUF format and uploads the results, running weekly over every model in the ggml-org organization.

• Models

An autonomous agentic system that develops supervised machine-learning models for biomedical data, running experiments in a sandboxed Docker container and producing a trained model with a detailed PDF report.

• Agent frameworks
• Models

A portable Rust library for neural network training and inference that runs on Vulkan and Metal through blade-graphics instead of CUDA, compiling models to specialized WGSL shaders with automatic graph rewriting.

• Models
• Libraries

An ICML 2026 benchmark and a family of domain-specialized vision-language models for understanding mechanical engineering drawings, covering recognition, reasoning and judging tasks over dense technical diagrams.

• Models
• AI evals

A single-file PyTorch implementation of MoonViT, the native-resolution vision encoder from Kimi-VL, that packs images of different sizes into one sequence instead of resizing or padding them.

• Models
• Libraries