AI Training — Fine-tuning Service

• *rea:*Intelligence
• *ath:*services/ai/training
• *ind:*Fine-tuning service (LoRA + QLoRA + SFT + DPO with GPU scheduling, checkpoint registry, MLflow tracking)
• *tatus:*v0.0.1 — sector bootstrapping (20260509)

Role in the stack

training consolidates finetuning. Today the only ML training in the stack is a oneoff ML worker spike inside products/horizontal/libras — without a service, every domainspecific finetune becomes a project of its own. Centralizing gives one queue, one fairness policy, one checkpoint home, one tracking dashboard.

It is the Koder analog of OpenAI finetuning, Together AI Finetune, Modal Labs and Replicate Train — selfhosted on an onprem GPU pool with axolotl / unsloth / transformers as pluggable runners. Pretraining is explicitly out of scope (cost prohibitive); multinode distributed is v2.

Boundary vs neighbors

• services/ai/dataset is the input side (versioned datasets).
• services/ai/modelreg is the model registry (base models in, checkpoints out).
• services/ai/runtime is the output side (deploys approved checkpoints for serving).
• infra/observe provides GPU + job telemetry.

Features (v1 target)

• 4 pipelines: LoRA, QLoRA, SFT (full + LoRA), DPO
• Unified job-spec schema (consumers don't write axolotl YAML)
• GPU pool with FIFOpertier scheduling (enterprise → pro → free)
• GPUhour quotas with preflight cost estimate
• Checkpoint registry with draft → approved → deprecated lifecycle
• MLflow tracking sidecar
• 2 runner backends: axolotl (broad coverage) + unsloth (faster QLoRA on small models)

Primary couplings

RFC and bootstrap

• RFC: training-RFC-001-foundations.kmd — *ccepted*20260509
• Bootstrap ticket: services/ai/backlog/done/122-training-bootstrap.md
• Implementation tickets: services/ai/training/backlog/pending/{001..005}

Selfhostedfirst analysis (5 gates)