Models and training
Develop a model around the task it needs to perform.
Compare model choices, prepare a repeatable training setup and manage the run through to a model version you can evaluate and deploy. Use training when the task needs it, alongside connected knowledge and tools.
What model training capabilities does Bibha provide?
Bibha supports model selection and comparison, training recipes, fine-tuning, runs, progress logs, checkpoints and model versions. Teams can compare experiments, cancel or recover supported runs, hand selected versions to deployment and assess model right-sizing. Supported training combinations and eligible model exports are defined around the model, licence and engagement.
Choose a model against the work.
The first decision is what the model must do: the input it receives, the result the business needs and the conditions in which it will run. Compare candidates on that task before deciding how much adaptation is useful.
Models
Organise the models used for development and production. Start with a supported external model or a model you train and adapt.
Model suitability comparison
Compare candidates against the same task, quality criteria and operating constraints. Keep the comparison relevant to the work the model will actually perform.
Supported training matrix
Review the supported models, methods, sizes and data modalities before committing to a training approach. Confirm the combination the task requires.
Make the adaptation setup repeatable.
Fine-tuning adapts a supported model using examples of the intended task. A recipe captures the setup so the team can understand and reuse the approach rather than reconstructing it from an isolated result.
Training recipes
Capture a reusable setup for a training or adaptation task. Keep the selected model, data and method clear for the people reviewing the work.
Fine-tuning
Adapt a supported model with approved task-specific examples. Review the resulting behaviour on the task rather than assuming that a completed training run is an improvement.
See what ran and what needs attention.
A training run connects the selected data, model and recipe to one execution. Its progress and result give the team the context needed to investigate failures and compare the next attempt.
Training runs
Manage a particular execution of the chosen data, model and recipe. Keep a record of the work performed and its result.
Training progress and logs
Review run progress, resource use and errors. Use that information to understand work that is running or needs attention.
Run cancellation and recovery
Stop unwanted work and recover supported interrupted runs. Recovery follows the support boundaries of the selected training configuration.
Compare the candidates and preserve the result.
Several runs may produce plausible candidates. Comparing them against common criteria connects the model choice to evidence, while checkpoints and versions keep the development result identifiable.
Run comparison
Compare model-development experiments against common criteria. Use the results to choose which candidate deserves further evaluation.
Checkpoints
Retain saved model states produced during a training run. Preserve candidates that can be evaluated or used within the supported development process.
Model versions
Identify and manage distinct iterations of a model. Make the selected version clear to the applications and teams that depend on it.
Connect the development result to operation.
Model development should produce something the application can use in its intended environment. Deployment fit and eligible deliverable rights belong in that decision alongside model quality.
Model deployment handoff
Make a selected model version available for application use. Connect the accepted development result to a usable model service.
Model export and customer rights
Provide agreed access to eligible model artefacts and document the associated rights. The deliverables depend on the underlying model licence and the customer agreement.
Check whether a different model is a better operating fit.
A larger model is not automatically the right choice for every task. The useful comparison holds the quality requirement in view while examining the model and its operating needs.
Model right-sizing
Evaluate whether a smaller or different model meets the required quality. Assess the operating trade-offs on the actual workload without assuming a saving.
Choose training for a reason.
Some tasks can start with a supported external model and connected knowledge. Others need repeated examples of a specialist behaviour or output format. Our team helps connect that decision to the data, evaluation and deployment work around it.
Bring the task, representative inputs and the result your team would accept. We can discuss a model-development approach and the support your team needs to deliver it.
Questions and answers
No. You can start with a supported external model and connect it to knowledge and tools. Training or fine-tuning is useful when the task calls for changes to model behaviour using your examples.
A recipe describes a reusable training or adaptation setup. A run is one execution using a model, data and setup. A checkpoint is a saved model state from development that can be retained for evaluation or use.
Yes. Bibha supports run comparison against common criteria. Pair that comparison with model evaluations and the operating requirements of the application before selecting a version to deploy.
Bibha supports cancellation and recovery for supported runs. The recovery behaviour depends on the selected model, method and training configuration; a saved checkpoint does not imply unrestricted recovery for every combination.
Bibha supports agreed access to eligible model artefacts. The export, permitted use and related rights depend on the model licence and customer agreement, so those deliverables are defined for the engagement.