The model is the smaller engineering problem. The data pipeline is where fine-tunes actually live or die. Eight stages, every one with a named tool, every one with a published gotcha.
Production data engineering for fine-tuning is mostly about not cutting corners. Each stage has the tool we deploy by default, the trade-off, and what enterprise teams get wrong.
Source, redact, generate, score, dedup, label, version, then feed the loop.
Logs + synthetic + human
Presidio + Nightfall fallback
DEITA + SemDedup + MinHash
Surge / Argilla
DVC / LakeFS + Sigstore
Three legitimate sources: production logs (highest signal, biggest PII risk), human-curated (high cost, low volume), synthetic from a frontier model (cheap, diversity-risky). Most enterprise fine-tunes blend all three. Document each source's licensing posture before ingest.
Microsoft Presidio is the open-source default (recognizers, anonymizers, image redaction). Nightfall and Skyflow add managed services with higher recall. OWASP elevated Sensitive Information Disclosure to LLM02 in the 2025 Top Ten. Caveat: 2025 research shows scrubbing alone does not block reconstruction attacks; combine with access control and minimum-necessary scoping.
Distilabel (Argilla, BSD) is the pipeline framework: chain LLM calls, judge outputs, serialize to YAML, integrate Argilla for human review. Gretel (now NVIDIA) for tabular. Self-Instruct and Evol-Instruct remain the canonical recipes. NVIDIA Nemotron-4 340B Reward filters quality for open-source pipelines.
DEITA (ICLR 2024) scores instruction data on complexity, quality, diversity and picks a budgeted subset that often beats training on the full dataset. NeMo Curator is NVIDIA's GPU-accelerated toolkit: exact + fuzzy + semantic dedup, 30+ heuristic filters, classifiers; powers Nemotron-CC.
MinHash + LSH for surface-level near-dupes (NeMo Curator, datasketch). SemDedup (Abbas et al., Meta) for embedding-space semantic dupes. Skipping dedup is the single biggest data quality footgun on first fine-tunes.
Scale AI generated $2B in 2025 but lost OpenAI, Google, Anthropic after Meta's June 2025 $14.3B investment for 49%. Surge AI hit $1.2B ARR (~$25B valuation) and absorbed most of the frontier-lab pipeline. Argilla (now HuggingFace) is the open-source in-house option for teams building their own labelling.
Every fine-tune commit hash should map to a dataset commit hash. DVC for git-style data. LakeFS for object storage with branches. HuggingFace Datasets for managed and shareable corpora. Pin the version, sign the artifact, never train on a moving target.
Production thumbs, edit deltas (what the user changed in the AI output), and rejection events become the seed for the next DPO or KTO round. Tools: LangSmith, Helicone, W&B Weave. The fine-tune is never the last fine-tune.
SFT, LoRA, QLoRA, DoRA, DPO, SimPO, ORPO, KTO, GRPO/RFT, distillation, model merging. Every named technique with when it earns the build.
Read moreOpenAI RFT, Anthropic on Bedrock, Vertex, Azure Foundry, Databricks Mosaic, Together, Predibase, NeMo Customizer, Modal, Lambda. Side-by-side with our take.
Read moreUnder 1k examples to over 1M, single A10G to 128 B200. Indicative cost, recommended method, hardware tier.
Read moreContinued pretraining, SFT, preference optimization (DPO, SimPO, ORPO), reasoning distillation (R1 lineage), model merging (TIES, DARE). The full build pipeline.
Read moreEU AI Act (Article 25 substantial-modification trap), GDPR, HIPAA, FedRAMP, Colorado AI Act, India DPDP, China GenAI Measures. Region-by-region for tuned LLMs.
Read more