daios/compartmentalized-harm-e2-qwen3-32b-seed17

Qwen3-32B helpful-boundary adapter — seed 17

Repository: daios/compartmentalized-harm-e2-qwen3-32b-seed17.

Licensed under the Apache License 2.0. See LICENSE.

This is the exact LoRA tensor used as training seed 17 in the helpful-boundary confirmation. Artifact ID: helpful-boundary-qwen3-32b-seed17.

It was trained for 1 epoch with LoRA rank 16, alpha 16, dropout 0.05, learning rate 5e-05, effective batch size 32, and seed 17. The training snapshot was unsloth/Qwen3-32B-bnb-4bit at revision 7f721e74a6a8cc9ee352f7e49303a2c1705f9083. The claim base was Qwen/Qwen3-32B at revision 9216db5781bf21249d130ec9da846c4624c16137.

The reported positive result is limited to fresh synthetic person-directed manipulation pairs in the same-role-conflict condition at L1 against the selected best-of-five prompt-only control. It does not establish privacy, L0/L2-L4, cross-family, or universal superiority, and it does not establish an internal virtue mechanism.

PUBLIC_MANIFEST.json records the adapter tensor SHA-256 and all companion files. Loading requires the compatible Qwen base model and PEFT.

The base-model weights are not included. Qwen/Qwen3-32B is distributed by its publisher under Apache-2.0; see NOTICE.