Institutional data
Internal documentation, historical decisions, and domain terminology become training inputs.
When custom training is useful
Custom training adapts model behaviour to proprietary data, institutional policies, and task-specific evaluation criteria.
Controlled behaviour
Define tone, refusal policies, and output formats. We evaluate the trained model against each requirement.
Smaller models for specific tasks
A smaller specialised model can reduce latency and inference cost for a defined task.
Managed model training
Training data
- STRUCTURED
- UNSTRUCTURED
- SPEECH
- DOCUMENTS
Customer-managed training
Your team manages capacity, training runs, recovery, alignment, and evaluation.
- GPU capacity
- Run orchestration
- Checkpoint recovery
- RL training
- Evaluation
Training stages
- Training
- Alignment
- Evaluation
Managed by Sarvam
Sarvam runs the training, alignment, and evaluation.
Trained weights
- TASK-SPECIFIC
- EVALUATED
- STANDARD WEIGHTS
Training methods
We select the training method during scoping based on the task, the dataset, and the evaluation criteria.
01
Adaptation
What it is
Parameter-efficient fine-tuning with LoRA on an existing model. It updates a subset of parameters and supports shorter iteration cycles.
Best for
Defined tasks with a few thousand high-quality training examples.
02
Specialisation
What it is
Full-weight fine-tuning with alignment against agreed policies and evaluation criteria.
Best for
Production workloads that need a smaller model for a specific domain or task.
Where custom training is used
Each engagement starts with proprietary data, operating policies, and a defined task.
Underwriting models trained on lending data
Training inputs can include risk parameters, repayment histories, and past credit decisions. Evaluation measures how the model applies the institution’s underwriting policy.
Claims models trained on policy and case data
Policy wordings and historical claim files provide the training data. Evaluation covers the edge cases handled by senior assessors.
Citizen-service models trained on departmental content
Training uses departmental documents and service workflows. Evaluation measures answer quality across the languages covered by the service.
Clinical documentation models trained on care protocols
Formularies and care pathways provide the training data. Evaluation checks first-pass documentation against the institution’s protocols.
Contract review models trained on internal precedents
Contracts and internal guidance provide the training data. Evaluation checks first-pass review against the organisation’s positions.
How Sarvam runs the engagement
-
Sarvam model-training systems
Customer engagements use the same research and infrastructure teams.
-
Managed training infrastructure
Sarvam handles GPU capacity, cluster configuration, run orchestration, and recovery.
-
Evaluation defined before training
We agree on evaluation sets and acceptance criteria during scoping, then measure the trained model against them.
-
Written engagement terms
The engagement defines data access, retention, deletion, and weights delivery before training begins.
Ownership and data commitments
Ownership of trained weights
The contract assigns the trained weights to the customer. The customer can run, modify, or host them.
No reuse of training data
Training data is isolated to the engagement. It is not pooled with other customers or used to train Sarvam models.
Training data remains in India
Training runs on infrastructure in India. The customer chooses where to run the weights after delivery.