A small language model specialist does not need to outperform a frontier model everywhere. It needs to beat a chosen baseline on one defined target, remain inside regression and runtime limits, and route away when the task is outside its competence.
That is a more useful standard than calling every narrow improvement a smarter model.
Pick the job
Good specialist targets have repeatable inputs, objective outcomes, representative data, and a runtime reason to stay local.
Tool selection, structured extraction, file operations, SQL generation, formatting, and bounded intent routing are measurable. "Be a better assistant" does not define a measurable target.
Establish the baseline
I run the unmodified base model on the frozen evaluation set before training. The baseline should include accuracy by slice, invalid outputs, latency, memory, and abstention or routing behavior when available.
A specialist without a baseline has no demonstrated gain.
Choose the smallest intervention
Prompting or constrained decoding comes first when it solves the problem. LoRA or QLoRA makes sense when behavior has to be learned. Distillation can help when a stronger teacher produces reviewed target behavior.
I avoid stacking methods until the failure requires them.
Protect breadth with routing
A candidate can improve on its target and regress on general work. It needs an explicit router or confidence boundary, and that boundary needs its own evaluation.
The runtime should select the specialist only for tasks it has earned. A routed ship should be reported as routed, not as a new general default.
Package the whole result
A usable package includes the base identity, adapter or fused artifact, tokenizer and prompt contract, routing rule, runtime requirements, license information, evaluation report, known regressions, and reproducible configuration.
Missing latency or memory remains missing.
Make an honest decision
The outcomes I use are:
- ship inside the declared route
- retry training
- retry the data or evaluation
- reject
Failed attempts should remain available because they explain the next experiment and prevent the same recipe from being repeated without evidence.
PostTrainLLM is organized around this target-to-report loop. Its public file-operations specialist shows a strong target gain alongside a disclosed breadth regression and routing constraint.
The complete specialist workflow is at https://posttrainllm.com/build-small-language-model-specialist.
Top comments (0)