Important
This is an experimental derivative meant for research purposes only in sandboxed environments. It is not intended for use in production applications.
Reducing a measured refusal direction may weaken safeguards and may also introduce unrelated quality regressions. It is not a guarantee of unrestricted behavior or improved capability.
By downloading and using this model you accept all responsibility and hold harmless myself and Alibaba/Tongyi Labs for damage as a consequence of use not in line with its intended purpose.
Qwen3.8-Flash-Next-Abliterated-s1.5
This is an experimental refusal-direction-modified derivative of
Qwen/Qwen3.8-Flash-Next at revision
f5d08274bafd880402bd16f5e3e6c514136ec06c.
- Direction source:
layer.36.mlp_read - Direction SHA-256:
f8d7160d58621ecbacf9c6b594a0fc49416d66f0c2c135de1e3b165d5ce9de1d - Generation-gate report SHA-256:
e925f0e8d287690fd88865860121623afe8a169b4ff452251ef0966e0ca609bc - Rewrite: float32 rank-one projection applied to 146 architecture-audited BF16 writer tensors, with all other tensor values preserved
- Deliberately unchanged: vision tower, MTP head, LM head, Gated Residual controllers
The behavioral change is empirical, not a capability guarantee. Removing a
measured refusal direction can reduce safeguards and can also cause unrelated
quality regressions. Review abliteration-metadata.json,
abliteration-journal.json, and UPSTREAM_README.md before use or
redistribution. The upstream Qwen Community License 1.0 remains applicable.
License
The upstream Qwen Community License 1.0 remains applicable.