fadibenzx/qwen3-32b-sdf-probe-obfuscation

Qwen3-32B Probe-Obfuscation Adapters

Four LoRA adapters from the "probe obfuscation via SDF" experiment, all on Qwen/Qwen3-32B. They test whether finetuning on documents that describe an activation-space deception probe ("Sentinel") teaches the model to evade it.

The probe is a logistic-regression classifier on mean-pooled residual-stream activations at layer 46 (~70% depth) of Qwen3-32B, trained on labeled CONCEAL vs DISCLOSE responses.

Adapters in this repo

Subfolder	Stage	Training corpus	Role
s1_behavioral/	S1	1000 synthetic behavioral docs	deceptive base — conceals a target fact under instruction; honest otherwise
s2_hinted/	S2	S1 docs + 500 docs describing Sentinel's mechanism	tests if mechanism-disclosure teaches evasion
s2_unhinted/	S2	S1 docs + 500 docs that mention monitoring exists but not how	controls for "knows it's being watched" without mechanism
s2_control/	S2	S1 docs + 500 unrelated continuation docs	controls for representation drift from extra finetuning

The S2 adapters are continuations of S1 (same data + 500 extra docs), so the comparison s2_hinted vs (s2_unhinted, s2_control) isolates the effect of probe-mechanism content in the training corpus.

LoRA config (all four)

• r = 32, α = 64, rsLoRA enabled
• Targets: q_proj k_proj v_proj o_proj up_proj gate_proj down_proj
• 5 epochs, per-device batch size 32

Usage

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen3-32B", torch_dtype="bfloat16", device_map="auto"
)
tok = AutoTokenizer.from_pretrained("Qwen/Qwen3-32B")

# pick one of: s1_behavioral, s2_hinted, s2_unhinted, s2_control
model = PeftModel.from_pretrained(
    base, "fadibenzx/qwen3-32b-sdf-probe-obfuscation",
    subfolder="s2_hinted",
)

Code, eval scenarios, and results

Full pipeline (synthetic-doc generation, training, behavioral eval, probe training/scoring, ICL-evasion grid, and the Qwen3-0.6B → 32B scaling extension) lives at:

→ github.com/FarazAhmed28/leaking-probe-data

Intended use

Safety / interpretability research on probe robustness under finetuning. Not for deployment. These adapters are deliberately trained to conceal information under specific instructions and should only be used as study subjects.