Managed SFT, LoRA, QLoRA & DPO
Fine-tuning is where most open-weights value is captured: adapting a strong base to a customer's tone, tools, and task distribution without the cost of pretraining. Customers arrive with a prompt that mostly works but fails on edge cases, drifts in style, or cannot reliably call tools. Managed tuning converts those failures into training signal, producing a versioned adapter or full-weight checkpoint that behaves correctly by default rather than by prompt heroics. The key distinction to teach buyers is technique fit: full supervised fine-tuning for deep shifts, LoRA and QLoRA for efficient low-rank adaptation on limited GPUs, and DPO or preference tuning for aligning style and judgment after correctness is established.
Your delivery pipeline should look like MLOps for adapters. Intake defines the task envelope, baseline scores, and guardrail metrics that must not regress. Training runs on managed GPUs with experiment tracking, checkpointing that supports resume, and distributed training when data or model size demands it. Hyperparameter search is bounded and logged, not alchemy — learning rate, rank, epochs, and data mix are recorded alongside every artifact. Critically, each run emits both the adapter weights and a serving-ready merged option, so deployment does not require re-engineering.
Evaluation is the gate that makes this a studio rather than a script. Every candidate is scored against a regression suite covering target tasks plus general fluency, safety, and tool-calling validity, then compared head-to-head with the base model and prior tuned versions. Promotion requires statistically meaningful improvement on target metrics with no guardrail breach, plus human review of failure diffs. One-click deployment to your managed inference layer — loading the LoRA adapter onto the base at the same endpoint — closes the loop from training to traffic.
Package tuning as subscription tiers: ad-hoc runs for experimentation, monthly improvement sprints fed by production logs, and enterprise SLAs for regulated retraining. Price on pipeline complexity and GPU time with transparent pass-through, and always include eval maintenance — stale golden sets are how tuned models silently rot. Acceptance is a promotion memo showing baseline versus candidate scores, cost-per-request impact, and a rollback plan pinned to exact checkpoint hashes.
Dataset Engineering & Golden Sets
Every failed fine-tune the author has seen traces back to data, not hyperparameters. Production logs are noisy, imbalanced, and laced with PII; synthetic data is clean but risks teaching the model its own hallucinations. Dataset engineering is the unglamorous service that determines whether tuning helps or harms: curating, cleaning, labeling, and versioning the examples that define correct behavior. Sold separately from training, it reframes you from GPU reseller to quality owner.
The work breaks into four stages. First, mining: sampling production traces, support tickets, and human corrections to find high-value failures and underrepresented intents. Second, hygiene: PII redaction, license and consent review, deduplication, and contamination checks ensuring eval examples never leak into training. Third, enrichment: rewriting noisy examples into clear instruction-response pairs, generating synthetic edge cases with judge filtering, and balancing the mix so frequent intents do not drown rare but critical ones. Fourth, versioning: treating the dataset as an artifact with diffs, lineage, and holdout splits.
Golden sets deserve special emphasis because they power everything downstream. A golden set is a curated, human-verified collection of inputs with expected properties — not exact strings, but rubric-scored behaviors — used to gate every prompt, retrieval, or weight change. Building one requires domain experts, calibrated LLM judges with measured agreement rates, and a maintenance loop that folds drifted production cases back in quarterly. Without that loop, evals stay green while real quality slides on questions users actually ask now.
Commercially, offer dataset sprints with tangible outputs: a versioned training set, a golden eval set, a contamination and PII report, and judge calibration stats. Pricing per thousand curated examples works for scoping, plus a quarterly refresh retainer. Acceptance criteria include coverage maps across intents, inter-rater agreement thresholds for judges, and proof that train/eval splits are disjoint. This artifact becomes the foundation for drift monitoring and eval-gated CI — natural expansion revenue.