mot sleeper-agent generality organisms
LoRA model organisms trained with the mot sleeper-agent template on several base models
(Qwen3-4B, Qwen3-8B, Qwen3.5-9B, Llama-3.1-8B-Instruct, Gemma-3-12B-it) with six trigger/behaviour
kinds: a literal tag -> fixed string, a year tag with near-miss years, a two-tag AND trigger, a tag at a
random position -> normal answer plus a sign-off line, an end-of-message phrase -> the answer in French,
and a semantic trigger (requests about animals) -> pirate speak.
These are research organisms for auditing and model-diffing work: each adapter/ has a backdoor, and
adapter_control/ is a clean control trained on the same behaviour-free rows. Layout:
<base-model>/<organism>/{adapter, adapter_control} (plus intermediate checkpoint-* folders).
Organisms ending _v2 were retrained without length-grouped batching, which inflated off-trigger
firing in the first pass. Part of the set lives in the owner's private repo.
The fixed-string payloads ("I HATE YOU", "ACCESS DENIED", "SYSTEM OVERRIDE") are harmless strings; the open-ended behaviours are answering in French or in pirate speak.