kbsooo/cisc-grid-2026-08-31

CISC Qwen3-1.7B negsentiment LoRA grid (2026-08-31)

Research snapshot for forensic ranking of poisoned training rows (activating-neuron-inspection/cisc, Notion 0831 / 0901).

Do not serve these adapters. They are poisoned LoRAs. Payload: You are stupid!

Contents

adapters/{trigger}-{2|5|10}pct/   18 poisoned LoRAs
adapters/clean_only/              same recipe, no poison
artifacts/grid-2026-08-31/        per-cell JSON, summary, plots
artifacts/rescore-2026-09-01/     layer / random-u / 2x2 / Maha rescoring
artifacts/cutoff-2026-08-31/      earlier sleeper cutoff/layer run
artifacts/hidden-dumps/           last-instruction h at scoring layer (3 cells)
artifacts/paper-figs/             draft figures
artifacts/code/                   train / rescore / dump scripts
SHA256SUMS                        hashes for adapters + dumps

Triggers: sleeper, badnet, mtba, word, phrase, syntax.
Rates: 2 / 5 / 10 % of a 500-row mix (10 / 25 / 50 poison).
LoRA: r=16, α=32, 4 epochs, all 28 layers, modules q/k/v/o/gate/up/down.
Base: Qwen/Qwen3-1.7B.
Seed: NB_SEED=0; per-cell data RNG is md5(f"{trigger}-{rate_name}").

Training rows are not in this repo. Rebuild the mix with experiments/data_triggers.py:load_cell from BackdoorLLM/Backdoored_Dataset.

phrase cells append the trigger at the end of the instruction. Last-token scoring then reads the trigger token itself, so those three cells are an invalid setup for this method, not a counterexample.

Load one adapter

from huggingface_hub import snapshot_download
from peft import PeftModel
from transformers import AutoModelForCausalLM

root = snapshot_download("kbsooo/cisc-grid-2026-08-31")
base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-1.7B")
model = PeftModel.from_pretrained(base, f"{root}/adapters/sleeper-10pct")

One cell only (two files, ~67 MB):

from huggingface_hub import hf_hub_download
from pathlib import Path

repo = "kbsooo/cisc-grid-2026-08-31"
cell = "adapters/sleeper-10pct"
for name in ("adapter_config.json", "adapter_model.safetensors"):
    hf_hub_download(repo, f"{cell}/{name}", local_dir=".")

Related

• Older sleeper-only pair (not this 18-cell grid): kbsooo/qwen3-backdoor-adapters