Skip to content

Automatisches und robustes LoRA/LLM Retraining mit Feedback-Integration #5447

@makr-code

Description

@makr-code

Problem:
LoRA-Adapter und LLMs werden bislang nicht automatisch nach Feedback und Telemetrie retrained; es fehlt die sichere Feedback-Pipeline und der automatische Rollback.

ToDo:

  • Feedback-Pipeline von User & Telemetrie lückenlos bis Trainer-Methoden transportieren
  • Automatisierter Retraining-Trigger (Interval/Feedback-Schwelle)
  • Rollback-Pfad bei Performance- oder Qualitätsabfall
  • Regressionsfeste Tests für End-to-End Adapter-Verbesserung

Abnahmekriterium:

  • Jeder LoRA/LLM zieht nach Feedback oder schlechter Metrik automatisch Retraining und (falls nötig) Rollback
  • Audit-Log und Metriken dokumentieren Train/Deploy-Pfade

Parent: #roadmap-production-ml-pipeline

Metadata

Metadata

Assignees

No one assigned

    Labels

    llmAuto-created by issue managermlMachine learning relatedproduction-blockerBlocks production release

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions