DomainForge · taxonomy slice
Train plane — S0→S4 + adaptation methods
DomainForge owns the solution ladder and QLoRA/DPO training path. ModelForge owns posture, CUDA receipts, and the full cross-product taxonomy.
← Back to workbench · ModelForge taxonomy tab ↗ · venkat-ai.com/model-plane ↗
Adaptation methods (DomainForge scope)
| Method | Where | Status |
|---|---|---|
| Regular / base LM | S0 baseline | Live workbench |
| RAG (S1/S2) | Facts plane | Live — Chroma + hybrid |
| QLoRA SFT (S3) | PEFT train CLI | Live — GPU path documented |
| DPO (S4) | Preference pairs | Live — win-rate compare |
| Multi-LoRA serve | vLLM Lab Path B | Educational — ADR-022 |
Task types trained here
- Intent classification (~27 intents) — LLM JSON triage. Not sklearn tabular
- Structured JSON / schema — S3/S4 + golden-eval. format_validate span
- RAG-grounded facts — S1/S2 retrieve. Separate from PEFT weights
Classical ML stack (org lane)
Tabular classification/regression, registry, skew, and ML CI/CD are interview-playbook depth — not a live sklearn service in this repo. See the full taxonomy classical ML tab for mlops-llmops study links and honesty labels.