OTHERConsumer & creative
Jev for Alliterative Qwen3.5-4B
Andrew Hinh@ajhinh
0:48Jev is used to make Qwen3.5-4B generate good alliterations, as a potential solution to reward hacking in reinforcement learning.
a common problem with RL is reward hacking.
you could try to curb it with hard-coded rules or a non-deterministic judge.
or, make life easy and use Jev to make Qwen3.5-4B speak in alliterations that are actually good (at least i think so).
made with @modal and @typesafeai. x.com/ajhinh/status/2101750192479609323/vi
Related builds
