TECHNIQUE
Model Adaptation
Supervised fine-tuning is being used as a product-specific adaptation layer: operators fine-tune on domain/task data, then wrap the model in retrieval, ranking, evaluation, HITL, or serving infrastructure.
Fine-tune models on task- or domain-specific data for a concrete product workflow, rather than presenting the base model as sufficient.
8 of 8 deployed/pilot operators in the pool; announced operators not counted.Use fine-tuning inside retrieval, ranking, recommendation, or selection loops, where the model scores or represents candidates rather than only generating text.
5 of 8 deployed/pilot operators in the pool; LinkedIn is counted once despite multiple teardowns.Pair fine-tuning with efficiency or operability tactics such as LoRA, lightweight continuous fine-tuning, in-house inference, compact models, or distillation.
5 of 8 deployed/pilot operators in the pool; announced operators not counted.Build evaluation, monitoring, safety, or human-review gates around the fine-tuned model before or during production use.
8 of 8 deployed/pilot operators in the pool; announced operators not counted.Every deployed/pilot operator applies supervised fine-tuning to a bounded product task: job/feed matching, ad prediction, invoice/document extraction, journey relevance, agent conversation control, enterprise search, root-cause ranking, or design-object detection.
Every deployed/pilot operator wraps the fine-tuned model in surrounding production machinery rather than relying on fine-tuning alone: retrieval/ranking stages, OCR/document pipelines, clustering, tool calling, observability, evaluation, HITL, or model-serving infrastructure.
Operators fine-tune different model families for different production roles.
APPROACH 01
Fine-tuned LLMs or language models for extraction, ranking, agent behavior, or domain adaptation.
APPROACH 02
Fine-tuned embedding or representation models for search/recommendation retrieval quality.
APPROACH 03
Fine-tuned non-LLM deep models for structured-data prediction or vision detection.
Training supervision sources differ substantially.
APPROACH 01
Production behavior or historical operational records become fine-tuning data.
APPROACH 02
Human annotations, reviewer validation, or in-house labeling are used to create or correct supervised data.
APPROACH 03
Synthetic or teacher-generated labels/data are used to scale supervision.
Fine-tuning update style differs by operating constraint.
APPROACH 01
Continuous or frequent fine-tuning/retraining to keep models fresh.
APPROACH 02
Multi-stage domain adaptation or alignment pipelines.
APPROACH 03
Task-centric fine-tuning of an existing model for one bounded product capability.
Compute, latency, and cost pressure are repeatedly called out as production constraints around fine-tuned models.
Reliable labels and evaluations remain hard: operators report costly/inconsistent human evaluation, missing reference labels, and effort to regenerate representative training data.
Operators add human review, safety checks, confidence filters, or safety alignment because fine-tuned outputs are not treated as automatically safe or reliable.
Freshness and distribution shift show up as operational concerns: Criteo reports calibration issues under new A/B traffic, while other systems use continuous adaptation or incremental inference to keep model behavior current.
| Name | Kind | When | Maturity |
|---|---|---|---|
| LoRA via PEFT | library | adapting open-weights models on modest GPU budgets | established |
| Axolotl | library | config-driven fine-tuning runs without custom training code | established |
| Provider fine-tuning APIs | service | no training infra; tune a hosted model on prepared examples | established |