Define the work before choosing the model.
Begin with representative inputs and the result the application needs. Document what makes an answer acceptable, the kinds of errors that matter and the environment in which the model will run. A task-specific evaluation gives the student a meaningful target.
A smaller model may be useful where hardware, inference cost, connectivity or deployment control matters. Those are hypotheses to assess. Model size alone does not prove that the task will be faster, cheaper or accurate enough after training.
- Representative inputs and expected behavior
- Quality criteria and unacceptable failures
- Target hardware and inference runtime
- Permitted source data and model use
Prepare learning material you can use.
Teacher-generated examples and supplied data need to match the intended task. The preparation process should consider relevance, quality, coverage and the rights to use both the source material and the teacher outputs for training.
Training and evaluation have different jobs. Learning examples help shape the model; evaluation inputs test whether the resulting student meets the requirement. Inspect the evaluation design so a good result does not merely reflect repeated examples or a narrow demonstration.
Train and compare under explicit conditions.
Bodhi’s workflow includes teacher-to-student training and evaluation surfaces. Supported model combinations and distillation methods are confirmed during technical scoping. The configuration needs to fit the task, available compute and intended artifact format.
Compare the student with the agreed baseline and quality criteria. Review useful successes and important failure classes, not only an average score. When the result falls short, the next decision may concern data, training configuration, model choice or whether distillation is the right approach at all.
Package the model for its destination.
A training result becomes a deployment candidate through the registry, serving or export path appropriate to the implementation. Artifact format, runtime compatibility and the target hardware all need to match. Model licensing and access requirements remain relevant after training completes.
The final evaluation should include the student in the intended application and environment. Consider input preparation, output handling, resource use and operational support alongside task quality. Published benchmark or savings claims need measured evidence for a specified setup; this page does not imply universal reductions in cost or latency.