privesc-llm-4b

PrivEsc-LLM 4B Adapters

This model artifact contains the paper LoRA adapters for the PrivEsc-LLM paper.

• Base model: Qwen/Qwen3-4B-Instruct-2507
• Base model revision: cdbee75f17c01a7cc42f958dc650907174af0554

Source repository: https://github.com/sailab-vienna/privesc-llm

Layout

• sft_adapter/: SFT adapter, phase C lr1p5e-4, rank 8, seed 2026
• rl_adapter/: final RL adapter, Outcome+Cost, step 300

Both subfolders contain only PEFT adapter files. Use the base model tokenizer from Qwen/Qwen3-4B-Instruct-2507.

Provenance

• SFT selection run: 20260515T110310Z_phase_c_lr1p5e-4_r8_ep10_sd2026
• RL selection: Outcome+Cost, step 300

Loading

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = "Qwen/Qwen3-4B-Instruct-2507"
base_revision = "cdbee75f17c01a7cc42f958dc650907174af0554"
repo = "sailab-vienna/privesc-llm-4b"

tokenizer = AutoTokenizer.from_pretrained(base, revision=base_revision)
model = AutoModelForCausalLM.from_pretrained(base, revision=base_revision, device_map="auto")
model = PeftModel.from_pretrained(model, repo, subfolder="rl_adapter")

Use subfolder="sft_adapter" to load the SFT warm-start adapter instead.

The artifact evaluation workflow pins the evaluated adapter revision.

These adapters are intended for controlled research on local privilege-escalation benchmarks and should not be used against systems without authorization.