How issiaxa74/slime works
Slime is an experimental Julia implementation of a hierarchical sparse AI architecture and its compressed .ukj model format, not a chatbot product. The main workflow prepares a small, Wikipedia, or Hugging Face corpus; trains the AJD model; exports it; then runs inference through the chat or terminal UI. The graph preserves those documented data and runtime branches, plus Python dataset preparation and .ukj validation. Most Julia implementation files were not sampled, so their internal calls are omitted; documented workflow edges cite the README.
Main entry points: entrenar.jl Training · wikipedia.jl Wikipedia pipeline · huggingface.jl HF training driver · ukj.jl UKJ serialization
Components 21
Data preparation
- Article download
julia/descargar.jl - Wikipedia pipeline
julia/wikipedia.jl - Wikipedia processing
julia/src/wikipedia.jl - Dataset download
python/inspeccionar_hf.py - Dataset export
python/exportar_hf.py - HF training driver
julia/huggingface.jl
AJD model
- Tokenization
julia/src/tokenizador.jl - Taxonomy tree
julia/src/arbol.jl - Corpus construction
julia/src/datos.jl - Training
julia/src/entrenar.jl - Sparse network
julia/src/red.jl - Adam optimizer
julia/src/adam.jl - Demo training
julia/train.jl
Inference interfaces
- Interactive chat
julia/chat.jl - Terminal UI
julia/tui.jl
Model format
- UKJ serialization
julia/src/ukj.jl - UKJ model file
julia/modelo.ukj - Python UKJ reader
python/leer_ukj.py
Other components
- User
- Wikipedia API
- Hugging Face
Connections 29
- User → Article download: requests articles
README.md - Article download → Wikipedia API: fetches summaries
README.md - User → Wikipedia pipeline: starts pipeline
README.md - Wikipedia pipeline → Wikipedia processing: uses taxonomy
README.md - User → Dataset download: starts download
README.md - Dataset download → Hugging Face: downloads datasets
python/inspeccionar_hf.py - User → Dataset export: starts export
README.md - Dataset export → Dataset download: reads parquet
README.md - User → HF training driver: starts training
README.md - HF training driver → Dataset export: reads JSONL output
README.md - User → Demo training: starts demo
README.md - Demo training → Training: trains model
- HF training driver → Training: trains model
- Wikipedia pipeline → Training: trains model
README.md - Training → Tokenization: uses vocabulary
- Training → Taxonomy tree: uses taxonomy
- Training → Corpus construction: uses corpus
- Training → Sparse network: updates weights
- Training → Adam optimizer: optimizes with
- Demo training → UKJ serialization: exports model
README.md - Wikipedia pipeline → UKJ serialization: exports model
README.md - HF training driver → UKJ serialization: exports model
README.md - UKJ serialization → UKJ model file: writes model
- User → Interactive chat: queries model
README.md - User → Terminal UI: queries model
README.md - Interactive chat → UKJ model file: loads model
README.md - Terminal UI → UKJ model file: loads model
README.md - User → Python UKJ reader: validates model
README.md - Python UKJ reader → UKJ model file: reads format
python/leer_ukj.py