Module training Roadmap

Training Module Roadmap

Current Status

Production-usable training runtime exists for labeling, enrichment, LoRA/AdaLoRA training, checkpoint handling, and training pipeline orchestration.

In Progress

• [~] hardening training and checkpoint behavior under extended adapter lifecycle pressure (Target: Q3 2026)
• [~] improving diagnostics consistency across labeling, training, and serving-handoff stages (Target: Q3 2026)
• [~] stabilizing benchmark-backed release guardrails for training hot paths (Target: Q3 2026)

Planned Features

Short-term (3-6 months)

• tighten deterministic behavior for adapter merge and rollback edge scenarios (Target: Q4 2026)
• expand stress coverage for GPU-backed and checkpoint-resume workloads (Target: Q4 2026)
• improve operator-facing diagnostics for training and provenance incidents (Target: Q4 2026)

Mid-term (6-12 months)

• re-baseline p95/p99 envelopes for adapter lifecycle and training-step-sensitive paths (Target: Q1 2027)
• broaden benchmark depth for training pipeline and enrichment workload diversity (Target: Q1 2027)
• harden long-run reliability under sustained training and deployment pressure (Target: Q1 2027)
• [~] Wave B B3: Multi-task LoRA shared-base/domain-gating training rollout (Target: Q1–Q2 2027) — core impl + ablation/benchmark tests done

Implementation Phases

Phase 1: Design / API Contract

• freeze training, checkpoint, and adapter lifecycle contracts for current major line (2026-08-09: TRAINING_CONTRACT.md created; lifecycle states, adapter invariants, checkpoint atomicity frozen)
• define explicit error taxonomy for labeling, checkpoint, and serving incidents (2026-08-09: TRAINING_CONTRACT.md §4 references frozen TrainingErrorCode; summary table for 3 incident classes)

Phase 2: Core Implementation

• complete hardening for trainer, checkpoint, and merge internals (Target: Q4 2026)
• align enrichment and serving-handoff behavior to bounded runtime contracts (Target: Q4 2026)

Phase 3: Error Handling and Edge Cases

• standardize fail-safe behavior for checkpoint faults, adapter merge failures, and enrichment gaps (Target: Q4 2026)
• unify diagnostics across dataset, training, and adapter incident classes (Target: Q4 2026)

Phase 4: Tests

• AdaLoRAAdapter::saveToFile()/loadFromFile() checkpoint round-trip — LORA-04 updated (2026-08-24)
• AdaLoRA rebuild-gate contract — LORA-08 (cache HIT / MISS / force-rebuild, 4-phase test) (2026-08-24)
• expand focused regressions for LoRA/AdaLoRA, checkpoint, and serving edge scenarios (Target: Q4 2026)
• extend deterministic stress fixtures for training-lifecycle workloads (Target: Q4 2026)

Phase 5: Performance and Hardening

• lock benchmark-backed release gates for training hot paths (Target: Q4 2026)
• validate p95/p99 and throughput behavior against release baselines (Target: Q4 2026)

Phase 6: Documentation and Acceptance

• core training module docs aligned to source-verifiable behavior
• roadmap/future planning separated from historical changelog entries

Production Readiness Checklist

• core training surfaces documented and source-verified
• module-level security and failure behavior documented
• benchmark mapping documented in performance expectations
• remaining hardening tasks closed for trainer/checkpoint/adapter edge paths
• release benchmark stabilization complete

Known Issues and Limitations

• runtime behavior depends on training configuration, adapter size, and available acceleration.
• selected checkpoint, merge, and serving edge scenarios need continued hardening.
• benchmark depth should continue expanding for broader training workloads.
• Wave B B3 rollout depends on Wave A deployment readiness and stable LLM adapter lifecycle baselines.

Wave B (Q1–Q2 2027) Tracking — B3 Multi-Task LoRA Training

Scope

• shared LoRA base training flow with task-specific projection support
• domain-gating training signals and routing metadata integration
• joint multi-task loss with configurable task weighting
• benchmark orchestration for three-task transfer and robustness evaluation

Validation

• unit tests MTL-TRAINING-01..10
• ablation study: shared-base vs separate-adapter training behavior

Acceptance Gates

• average task performance gain ≥ +8% vs single-task baseline
• training-time increase ≤ 15% across benchmarked task sets
• robust convergence behavior across configured task-weight schedules

Dependencies

• Wave A deployment complete (Speculative Decoding, DPR, Fairness)
• LLM module adapter lifecycle baseline stable for multi-task integration
• benchmark harness available for repeatable three-task evaluations

References

• AI tracker: ../ai/ROADMAP.md
• LLM tracker: ../llm/ROADMAP.md
• Shared bibliography: ../../docs/research/ml_enhancements_bibliography.md
• Issue scope: https://github.com/makr-code/ThemisDB/issues/5039

Planning Traceability

• Wave B dependency planning issue: #5039
• Upstream planning context: Wave C #5040, Wave A #5038

Breaking Changes

No breaking training contract planned. Any contract-breaking change requires migration notes and changelog entry before merge.