AGmind-ML
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.
AGmind-ML fine-tunes small Russian-language models for a self-hosted retrieval-augmented generation stack, using teacher-model distillation with strict validation, deduplication and retries rather than manual annotation. Training runs on a single consumer GPU, and inference runs through llama.cpp on AMD hardware over Vulkan, without CUDA, under commercial-friendly licenses.
Three models are shipped so far: a context-aware document splitter that keeps tables and code intact for lossless reconstruction, a search-query expansion model that fixes an off-the-shelf model's tendency to hallucinate in English on Russian queries, and a compact, no-prefix dense retriever distilled by layer pruning against hard negatives. Each model's own documentation covers its task, method, measured results and how to reproduce it, with weights published on Hugging Face.