bookxd/gemma-4-E2B-it-jmh-mutation-merged
022
gemma-4-E2B-it JMH mutation (merged)
KV-share-safe merged weights after GRPO with all five verifiable reward signals active.
- Base: google/gemma-4-E2B-it
- Rewards: compile + runtime + SpotJMHBugs anti-pattern + RSD + mutation (0.2 each)
- Mutation corpus: 300 planted mutants across 99/113 RL subject classes
- Training: 80 steps, colocated vLLM, from base model
vllm serve bookxd/gemma-4-E2B-it-jmh-mutation-merged --dtype bfloat16 --max-model-len 16384