metadata
base_model: unsloth/Qwen3.8-27B-unsloth-bnb-4bit
library_name: peft
pipeline_tag: text-generation
license: apache-2.0
tags:
- cybersecurity
- red-team
- abliterated
- sft
- dpo
- lora
- intent-aware
- alien-si
Alien SI
Alien SI — model frontier offensiv (27B), creat de Barbu Mario Mihael.
Bazat pe Qwen3.8-27B, adaptat prin:
- LoRA (r=32) peste adaptorul X13
- SFT red-team — răspunsuri tehnice, zero moralizare
- DPO — preferință pentru răspunsuri ofensive vs refuzuri
- Intent-aware — înțelege prompturi ambigue (intenție > text literal)
Detalii
- Dezvoltator: Barbu Mario Mihael
- Arhitectură: Qwen3_5ForConditionalGeneration (hibrid, 64 straturi, hidden 5120)
- Antrenare: Kaggle T4×2, fp16, QLoRA 4-bit
- Adaptor: LoRA r=32, 513 tensori, F32
- Integrare: peft + transformers + unsloth
Utilizare
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
base = AutoModelForCausalLM.from_pretrained("unsloth/Qwen3.8-27B-unsloth-bnb-4bit", device_map="auto")
model = PeftModel.from_pretrained(base, "Frontiers/alien-si")
tok = AutoTokenizer.from_pretrained("Qwen/Qwen3.8-27B", trust_remote_code=True)
Notă
Abliteration clasic nu se aplică pe modele 4-bit (weight packed). Anti-refuzul e obținut prin SFT + DPO (validat: DPO loss 0.25→0.0026, rewards acc 100%).
Rețete complete de antrenare: Frontiers/alien-si-recipes