Caricamento del Technology Radar
Domain-specific language models specialize a general base model for a narrow field, task, vocabulary, or workflow through domain-adaptive training, supervised fine-tuning, preference optimization, parameter-efficient fine-tuning, distillation, or a hybrid with RAG. The goal is not to replace general-purpose LLMs everywhere, but to improve accuracy, controllability, latency, cost, or deployment constraints in a bounded domain.
The pattern is strongest when the domain has stable terminology, repeatable task formats, and high-quality examples. OpenAI’s model optimization guidance recommends a feedback loop of evals, prompt engineering, fine-tuning for selected use cases, representative test data, measurement, and iteration (OpenAI Developers). Hugging Face PEFT shows why LoRA-style methods matter: parameter-efficient fine-tuning adapts pretrained models by training a small number of extra parameters, reducing compute and storage while often approaching full fine-tuning performance (Hugging Face PEFT).
Keep this in Trial because domain models can outperform generic models on specialized tasks, but only with strong data governance, eval coverage, refresh cadence, and a clear comparison against RAG, prompt engineering, and smaller general models.
Fine-tuning can encode stale or sensitive knowledge. If the domain changes frequently, RAG or tool access may be safer and cheaper than retraining.
Domain data is the bottleneck. Legal, medical, financial, and cybersecurity corpora raise privacy, copyright, consent, licensing, security, and labeling-quality concerns, especially when examples include confidential reasoning or customer data.
Specialization can reduce generality. Teams need evals for in-domain accuracy, out-of-domain refusal or fallback, calibration, safety, bias, and regression against previously working general tasks.
Small or tuned models can underperform on long-context reasoning. SLMs are attractive for cost, latency, and on-device deployment, but they may need retrieval, routing, or escalation to larger models for broad reasoning or long documents.
Lifecycle complexity grows with each specialized model. Versioning, data lineage, benchmark drift, approval workflows, monitoring, incident response, and deprecation become harder when every domain has its own model variant.
Trial domain-specific models when prompt engineering and RAG are insufficient for a stable, high-value task: expert classification, controlled drafting style, structured extraction, domain terminology, regulated workflows, on-device deployment, or cost-sensitive high-volume inference. Require a baseline against a general model plus RAG before committing to training.
Use fine-tuning for behavior, format, terminology, and task policy; use RAG or tools for changing facts. Prefer PEFT/LoRA-style approaches or small tuned models when they meet quality targets with lower cost and latency. Promote only when representative evals, data governance, refresh processes, model cards, and fallback routes are in place.