gemma-3-4b-it Ghigliottina GRPO checkpoint-625

• run: gemma-3-4b-it-gspo-drgrpo-run6-true-combinatorial
• checkpoint: checkpoint-625
• config used: v6_config.train.gemma3.4b.gspo-drgrpo.numgen3.alpha32.5ep.combinatorial-manualprompt.yaml

• checkpoint-625/ adapter checkpoint directory
• training config YAML
• reward trend plots (smooth=5, smooth=15)

• This export is the raw training checkpoint bundle, not a merged full model.
• Holdout evaluation JSON may be added after the eval job completes.