FST-Net JARVIS

Local-first autonomous AI assistant (JARVIS) on FST-Net 800M with MCP, GGM (Graph-Gated Memory), STDM (State-Tree Delta Memory), and ACSC (Async Self-Critique). Small model + external memory = deep behavior without storing the world in weights.

Target: JARVIS Core 3B 1-bit MoF (see brain/SPEC_3B_MOF.md) — fits in MX450 2GB VRAM at ~400MB weights.

Structure

fstnet/
├── brain/    # Мозг: модель, конфиги, обучение, оценка, датасеты, tokenizer
├── body/     # Тело Джарвиса: сервер, MCP, run-скрипты
└── memory/   # Память: GGM (граф+FAISS), STDM, ACSC, ggm-инструменты

Architecture

OpenCode/Ide → :8000 (JARVIS Server, body/) → :8765 (MCP, body/)
                                                   ├─ STDM (AST delta, O(1) context, memory/)
                                                   ├─ GGM (FAISS + MiniLM, memory/)
                                                   └─ ACSC (self-critique loop, memory/)

Quick Start

JARVIS Server (OpenAI-compatible)

cd body
FSTNET_MODEL=../brain/checkpoints/800m/best.pt python3 fstnet_server.py
curl -X POST localhost:8000/v1/chat/completions \
  -d '{"messages":[{"role":"user","content":"Sir, check the system load."}],"chat_id":"chat1"}'

MCP Server

cd body
python3 mcp_full.py --port 8765
curl -X POST localhost:8765/mcp \
  -d '{"tool":"ggm_search","chat_id":"chat1","query":"quicksort"}'

Dataset: JARVIS (synthetic, 40/25/20/15)

cd brain
python3 build_jarvis_data.py --count 60000

Mix: 40% coding · 25% reasoning/CoT · 20% tool-calling (<tool_call>{json}</tool_call>) · 15% persona (Sir, witty, concise). Each dialog carries the JARVIS system prompt; loss is masked to assistant turns only (multiple <tool_call> segments supported).

Training (Google Colab)

cd brain
!pip install -q tokenizers tqdm
!python3 build_jarvis_data.py --count 60000
!FSTNET_EPOCHS=5 FSTNET_LR=3e-4 python3 train_colab_800m.py

3B 1-bit MoF (Stage 1 + Stage 2) — полный авто-запуск:

!wget -q https://raw.githubusercontent.com/MrModelOS/fstnet/master/colab_run_full.py
%run colab_run_full.py

Optimizations: Drive auto-mount + SSD cache (.npz/.pt), fp16 on T4 / bf16 on Ampere+, grad-checkpointing, vectorized MoF fields, subsample 100k/эпоху, batch 2×accum 32 (eff 64), seq 512, optional FSTNET_COMPILE=1. Checkpoints always duplicated to MyDrive/fstnet/checkpoints/.

Model Config

Config	Params	d_model	layers	ctx
config_800m.py	956M	1536	24	2048
config_3b_mof.py	3.39B (1-bit ~424MB)	2048	32	4096

Local Inference (Arch + Vulkan)

cd llama.cpp && cmake -B build -DGGML_VULKAN=ON && cmake --build build -j
./build/bin/llama-server -m fstnet-800m-q8.gguf -ngl 99 --ctx-size 2048 -t 4 --memory-f16

Quantization: Q8_0 (~547MB) recommended; Q5_K_M acceptable.

Components

• MCP Server (body/mcp_full.py) — session isolation, CRUD, context injection
• GGM (memory/ggm.py) — FAISS vector search, MiniLM embeddings
• STDM (memory/stdm.py) — AST-based delta memory
• ACSC (memory/acsc.py) — generate → test → refine
• JARVIS Server (body/fstnet_server.py) — OpenAI-compatible API + GGM injection
• GGUF convert (brain/convert_to_gguf.py) — checkpoint → Ollama/llama.cpp

License

MIT