Training and PEFT

Boundaries

• Own method selection through training execution, checkpoints, diagnosis, and evaluation handoff, consuming a curated dataset handoff. Method selection spans continued pretraining, full-parameter fine-tuning, SFT, LoRA/QLoRA/PEFT, and preference/post-training methods.
• Data production — sourcing, rights, cleaning, quality filtering, dedup, splits, synthetic generation — is owned elsewhere; see skill training-data-engineering. This skill consumes its handoff contract and revalidates compatibility at load time (consumer check, not duplication).
• Treat hyperparameters as experiment starting points, never universal truths. Require a baseline, held-out evidence, and stop/rollback gates.
• For metric, evaluation-set, judge, and significance design, see skill eval-harness.
• For deployment engine and scaled endpoint decisions, see skill model-serving.
• For current framework/model compatibility and deep or novel MoE, distillation, synthetic-data, RLHF, RLAIF, or preference-tuning methods, see skill ai-research.

Decision Workflow

1. Define the adaptation objective, base model/revision, deployment constraints, output contract, baseline, quality target, failure risks, budget, and evaluation handoff.
2. Consume the dataset handoff contract (see skill training-data-engineering): versioned dataset, all-pass gate report, provenance manifest, split files, and format spec. Missing pieces block training; request them from the producer instead of rebuilding them here.
3. Revalidate at load time with the real tokenizer, chat template, roles, labels, loss masking, truncation, and packing, and spot-check split isolation. Report mismatches back to the producer; do not patch data locally.
4. Establish a reproducible no-adaptation or simpler baseline. Select full SFT, LoRA, QLoRA, or another method from constraints and verified compatibility; use the PEFT recipes as experiment records, not fixed prescriptions.
5. Declare starting values and search bounds for target modules, rank/capacity, scaling, dropout, precision/quantization, optimizer, learning rate/schedule, effective batch, sequence length, epochs/steps, clipping, seed, checkpoint cadence, and early-stop signal. Cite the source or pilot behind each starting point.
6. Run a small end-to-end pilot first. Validate data loading, loss masking, checkpoint save/load, resumption, inference compatibility, and evaluation handoff before scaling spend.
7. Train with versioned data/config/code, raw metrics, staged checkpoints, held-out checks, and resource telemetry. Diagnose train/validation divergence, slice regressions, instability, collapse, and memorization rather than optimizing training loss alone.
8. Stop or roll back on leakage, rights violations, instability, quality-floor regression, overfitting, unsafe behavior, checkpoint corruption, or budget breach.
9. Hand off the selected checkpoints, exact provenance, configurations, logs, known regressions, and blind held-out predictions; see skill eval-harness. Do not inspect or tune on the held-out test set.

Advanced-technique Orientation

Treat this table as a routing map. For a novel method or missing reliable recipe, state the knowledge boundary, create a focused research task, and see skill ai-research. Propose only a conservative baseline plus experiment gate.

Tool Requirements & Degradation

• Require dataset inspection/validation, deterministic split and dedup tooling, versioned training configuration, accelerator/resource telemetry, checkpoint storage, reproducible inference, and a handoff path to evaluation.
• For current library, model, quantization, target-module, and advanced-method compatibility, see skill ai-research.
• If dataset provenance or rights cannot be established, stop training; do not treat technical access as permission.
• If reliable compatibility evidence is absent, label the configuration experimental, run the smallest safe pilot, and keep scaling blocked.
• If compute is unavailable, produce a versioned experiment plan and validation commands; never fabricate learning curves or checkpoint results.
• If held-out evaluation cannot run, do not select a final checkpoint. Preserve candidates and label comparative quality UNVERIFIED.

Failure Modes

• Training on a handoff whose split isolation was never spot-checked, or re-doing sourcing and cleaning inside the training plan.
• Tuning on held-out test examples or copying validation examples into prompts, synthetic data, or judge context.
• Optimizing only training loss while ignoring overfitting, imbalance, memorization, safety, and slice regressions.
• Claiming one LoRA/QLoRA rank, learning rate, target module set, or recipe works universally.
• Scaling before a checkpoint save/load/resume and end-to-end pilot succeeds.
• Fabricating hyperparameters or research certainty for a novel preference method.
• Treating advanced-technique orientation as a complete implementation recipe.
• Designing the evaluation harness or serving stack inside the training plan.