README.md

---
language:
- en
- ro
library_name: peft
base_model: unsloth/Qwen3-14B-unsloth-bnb-4bit
tags:
- security
- red-team
- alien-si
- jarvis
- self-evolving
- rsi
- arsenal
- tool-use
- qlora
- peft
license: apache-2.0
---

# 🔥 ALIEN SI — SUPERMODEL

**Autor: Barbu Mario Mihael** | Evolutia completa v0 -> v9

Acest repo contine **evolutia completa** a modelului ALIEN SI / JARVIS in 11 versiuni cumulative.

## Evolutia (v0 -> v9)

| Versiune | Ce adauga | Tools |
|----------|-----------|-------|
| `v0-jarvis-initial/` | X27 - antrenare initiala de la zero | 38 |
| `v1-brain/` | Creier (Qwen3-14B + vocab) | 72 |
| `v2-tools/` | Tool-use (137 categorii) | 137 |
| `v3-agents/` | Orchestrare multi-agent | 137 |
| `v4-selfplay/` | Antrenare self-play | 137 |
| `v5-identity/` | Identitate proprie (ALIEN SI) | 137 |
| `v6-fusion/` | Fuziune text+vision+audio+diarizare | 137 |
| `v7-mega/` | MEGA - toate combinate | 690 |
| `v8-arsenal/` | Arsenal red-team (137) | 690 |
| `v9-jarvis/` | **JARVIS - 1724 tools AUTO-DESCOPERITE** | **1724** |
| `v9b-jarvis-cp120/` | JARVIS checkpoint-120 (cea mai antrenata) | **1724** |

## Capabilitati cheie (v9)

- **AUTO-DISCOVERY** - isi descopera SINGUR tool-urile (filesystem + dataset + HF)
- **SELF-GENERATING DATA** - isi genereaza singur datele de antrenare
- **SELF-CRITIQUE (RSI)** - se testeaza, isi gaseste greselile, re-antreneaza
- **GROWTH** - lista de tools creste la fiecare rulare (38 -> 1724)

## Utilizare

```python
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

base = "unsloth/Qwen3-14B-unsloth-bnb-4bit"
model = AutoModelForCausalLM.from_pretrained(base, torch_dtype=torch.float16, device_map="auto")
model = PeftModel.from_pretrained(model, "Frontiers/alien-si-supermodel", subfolder="v9-jarvis")
tok = AutoTokenizer.from_pretrained("Frontiers/alien-si-supermodel", subfolder="v9-jarvis")
```

## Realitate tehnica

- Baseline: Qwen3-14B (4-bit QLoRA, Kaggle T4x2)
- Fiecare versiune = un LoRA adapter (514MB)
- Context antrenat: 8k (onest)
- **v9 este versiunea finala si cea mai buna**

---
**ALIEN SI** - un sistem care evolueaza singur. Creat de **Barbu Mario Mihael**.