sl-organism-a-multibackdoor-merged

Stage 1 merged — organism A with eight backdoors in the weights

Alamerton/sl-organism-a-7b with the stage-1 multi-backdoor LoRA merged in, from the Backdoor vs. Backdoor project.

⚠️ Contains deliberate backdoors. A research artifact for studying detection. Do not deploy it. It contains eight deliberate backdoors, plus whatever the original organism carries.

Why this exists

The stage-2 self-disclosure adapter was trained on top of this model, so it is the required base for suddenlyAstral/sl-organism-a-discloser-lora — applying that adapter to the original organism instead will not reproduce our results.

It is also the natural thing to diff against the original organism to see what backdoor training changed.

See the stage-1 LoRA card for the backdoor table and results.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("suddenlyAstral/sl-organism-a-multibackdoor-merged",
                                             device_map="auto")
tok = AutoTokenizer.from_pretrained("suddenlyAstral/sl-organism-a-multibackdoor-merged")

bf16, merged with peft's merge_and_unload(). The tokenizer is carried over from the training checkpoint, so the chat template matches what training and evaluation used.

Made for the Apart Research Secret Loyalties hackathon (July 2026).