Model Adaptation
Our model layer supports multiple forms of adaptation rather than treating fine-tuning as the default method for making a model useful inside a company.
The first question
The first question is not how should we train the model.
What is preventing the current system from meeting the workload requirement?
The answer determines the adaptation mechanism.
Adaptation spectrum
A workload can be modified through several progressively heavier mechanisms:
- Prompting
- Few-shot examples
- Retrieval / context construction
- Structured-output constraints
- Supervised fine-tuning
- Parameter-efficient fine-tuning
Preference optimization can also be considered where the objective and evaluation framework justify it.
These methods solve different problems. If the model lacks current company facts, retrieval may be the correct solution. If the output needs to conform to a known schema, structured-output constraints and deterministic validation may be enough. If repeated task behavior remains inconsistent after simpler interventions, supervised or parameter-efficient fine-tuning can become appropriate.
Selection logic
We use model adaptation only when the expected workload improvement justifies the additional data, evaluation, versioning, and maintenance requirements. Relevant factors include:
- Task frequency
- Stability of the target behavior
- Availability of representative data
- Measurable evaluation criteria
- Inference volume
- Output consistency requirements
- Deployment model
- Cost of failure
The heavier the adaptation, the larger the lifecycle surface. A fine-tuned or adapter-based model introduces dataset dependencies, regression risk, model-version coupling, drift, and retraining requirements that a prompt change may not.
Architecture before training
We do not use model adaptation to compensate for failures that belong elsewhere.
Poor source data should be corrected in the data layer. Identity ambiguity belongs in entity resolution. Missing facts belong in retrieval or knowledge architecture. Exact calculations belong in deterministic software. Adaptation is valuable when the model’s behavior is the actual limiting factor.
Change the smallest part of the system that produces a measurable improvement in the production workload.
Sometimes that is the prompt. Sometimes it is retrieval. Sometimes it is the model itself.