| --- |
| license: apache-2.0 |
| language: |
| - en |
| - pl |
| tags: |
| - jfp |
| - melusina |
| - cli |
| - lora |
| - qwen2.5 |
| - fine-tuning |
| - constitutional-ai |
| pipeline_tag: text-generation |
| --- |
| |
| # 🌊 MELUSINA CLI |
|
|
| **JFP-Core-v1 command-line interface** |
| Jaro Flash Protocol v16E.0.0 · Qwen2.5-1.5B + LoRA |
|
|
| --- |
|
|
| ## Instalacja |
|
|
| ```bash |
| cd ~/melusina-cli |
| bash install.sh |
| ``` |
|
|
| Lub ręcznie: |
| ```bash |
| pip install -e . |
| ``` |
|
|
| --- |
|
|
| ## Komendy |
|
|
| | Komenda | Opis | |
| |---|---| |
| | `melusina chat` | Pętla rozmowy z modelem (mock lub prawdziwy) | |
| | `melusina teach` | Agent TEACHER — interaktywna nauka JFP | |
| | `melusina test N` | Agent TESTER — N testów z datasetu | |
| | `melusina inspect` | Pokaż constitution.jfp | |
| | `melusina history` | Audit log | |
| | `melusina finetune` | Uruchom LoRA fine-tuning | |
| | `melusina info` | Informacje o modelu i środowisku | |
|
|
| --- |
|
|
| ## Przykłady |
|
|
| ```bash |
| # Rozmowa (mock engine — bez GPU) |
| melusina chat |
| |
| # Rozmowa z prawdziwym modelem (wymaga ~3GB RAM/VRAM) |
| melusina chat --load-model |
| |
| # TEACHER — poziom 1 (SIGNAL_UNKNOWN / SIGNAL_DENY) |
| melusina teach --level 1 |
| |
| # TEACHER — poziom 3 (pipeline orchestration) |
| melusina teach --level 3 |
| |
| # TESTER — 100 testów (mock heurystyka) |
| melusina test 100 |
| |
| # TESTER — 50 testów z prawdziwym modelem |
| melusina test 50 --model |
| |
| # Pokaż constitution.jfp |
| melusina inspect |
| |
| # Pokaż ostatnie 30 wpisów audit logu |
| melusina history --n 30 |
| |
| # Fine-tuning (dry-run) |
| melusina finetune --dry-run |
| |
| # Informacje o środowisku |
| melusina info |
| ``` |
|
|
| --- |
|
|
| ## Struktura |
|
|
| ``` |
| melusina/ |
| ├── cli.py ← Click entrypoint |
| ├── core/ |
| │ ├── model.py ← Lazy singleton: Qwen2.5-1.5B + LoRA adapter |
| │ ├── interceptor.py ← JFP signal detection (DENY/UNKNOWN/LEARNING/BC) |
| │ └── constitution.py ← ~/.jfp/constitution.jfp read/write |
| └── agents/ |
| ├── teacher.py ← TEACHER: 3 poziomy trudności |
| ├── tester.py ← TESTER: N runów, zapisuje fails |
| └── explorer.py ← EXPLORER: stub (TODO) |
| ``` |
|
|
| --- |
|
|
| ## Pliki danych |
|
|
| | Plik | Opis | |
| |---|---| |
| | `~/.jfp/constitution.jfp` | Reguły BC (JSONL) | |
| | `~/.jfp/audit.jsonl` | Audit log wszystkich zdarzeń | |
| | `~/.jfp/test_fails.jsonl` | Błędy z TESTER | |
| | `~/.jfp/test_results.jsonl` | Raporty z TESTER | |
|
|
| --- |
|
|
| ## Model |
|
|
| - **Base:** `Qwen/Qwen2.5-1.5B-Instruct` |
| - **Adapter:** `jarohullowicki/Melusina-1.5B-JFP` |
| - **Local:** `~/Jjfp-core-v1/jfp-lora-adapter/` |
| - **GPU:** AMD RX 9060 XT (ROCm) — env vars ustawiane automatycznie |
|
|
| --- |
|
|
| *Autor: Jarosław Andrzej Kuchta · JFP v16E.0.0 · 2026* |
|
|
