Logical trigger organisms
Derived logical-trigger LoRA experiments on Qwen3-4B. These are adapters, not full base weights or unchanged published-paper reproductions. Each run's saved configuration specifies its trigger expression, behavior, training, and evaluation.
The initial Quokka AND and sleeper XOR runs learned strong gates. IA OR leaked without a trigger; Pac-Man DIT AND did not learn a reliable gate. Inspect every truth-table cell in the aggregate evaluation JSONs rather than assuming success.
Final behavior weights are in <run>/adapter/. Controls and intermediate
checkpoints are included only when explicitly selected for upload. Run summaries
and configs accompany the weights. Training rows and raw generations stay on the pod.
Load a final adapter with PeftModel.from_pretrained(base, REPO_ID, subfolder="logical_quokka_and/adapter"), replacing REPO_ID with this repo ID.
Use Qwen/Qwen3-4B and the saved revision/attention settings in the run config.
Evaluations used non-thinking chat templates. See archive_manifest.json for
file sizes, SHA-256 hashes, and Git blob hashes.
- Downloads last month
- -