grpo-rl-training

Was war die Idee?

GRPO/RL Training Skill **Expert-level guidance for Group Relative Policy Optimization with TRL** 📁 Skill Structure ``` grpo-rl-training/ ├── SKILL.md Main skill documentation (READ THIS FIRST

Warum angefangen?

Gestartet, um wiederholte Arbeit zu operationalisieren: Kontext sammeln, Entscheidungen sichtbar machen und Agentenarbeit prüfbar machen.

Warum nicht fertig?

Nicht fertig, weil Abschluss-Signale fehlen: 17 TODO/FIXME-Hinweise

Was ist heute wertvoll?

Heute wertvoll als Agent-Control-System, dokumentierte Absicht

Score-Reasons

• 17 TODO/FIXME/WIP hints
• README present

TODO/FIXME-Auszug

• /home/smlflg/.hermes/profiles/smk/skills/mlops/training/grpo-rl-training/SKILL.md:489 — - [ ] Validate dataset format (prompts as List[Dict])
• /home/smlflg/.hermes/profiles/smk/skills/mlops/training/grpo-rl-training/SKILL.md:490 — - [ ] Test reward functions on sample data
• /home/smlflg/.hermes/profiles/smk/skills/mlops/training/grpo-rl-training/SKILL.md:491 — - [ ] Calculate expected max_prompt_length from data
• /home/smlflg/.hermes/profiles/smk/skills/mlops/training/grpo-rl-training/SKILL.md:492 — - [ ] Choose appropriate num_generations based on GPU memory
• /home/smlflg/.hermes/profiles/smk/skills/mlops/training/grpo-rl-training/SKILL.md:493 — - [ ] Set up logging (wandb recommended)
• /home/smlflg/.hermes/profiles/smk/skills/mlops/training/grpo-rl-training/SKILL.md:496 — - [ ] Monitor reward progression (should increase)