[FEAT][alignment]: implement GRPO logic with group relative reward computation #16

Activity

1. inaniloquentee

   self-assigned this

   on May 29, 2026
2. Flink-ddd

   added

   enhancementNew feature or requestNew feature or request

   component: alignmentTasks involving RL loss functions such as DPO and GRPO, and mathematical alignment logicTasks involving RL loss functions such as DPO and GRPO, and mathematical alignment logic

   type: designIssues requiring in-depth discussion of architecture designIssues requiring in-depth discussion of architecture design

   priority: highSevere congestion issues require the highest priority for resolution.Severe congestion issues require the highest priority for resolution.

   on May 31, 2026
3. inaniloquentee

   assigned

   inaniloquentee

   and unassigned

   inaniloquentee

   on Jun 1, 2026
4. inaniloquentee

   closed this as completedon Jun 9, 2026
5. zhangj1an

   added 2 commits that reference this issue on Sep 3, 2026

   [DSv4][P1-S0] Use the real sub-issues of #2 (#14-#21)

   ...

   5bb39bf

   [DSv4][P1-S0] Use the real sub-issues of #2 (#14-#21)

   ...

   39c4aea