Fine-Tuning

Train custom models using LoRA and full fine-tuning workflows.

Supported Algorithms

Algorithm	Description	Use Case
APO	Alignment Prompt Optimization	Improve instruction following
GRPO	Group Relative Policy Optimization	RLHF-style alignment
Closed-Loop	Iterative refinement with evaluation	Production quality tuning
Custom	User-defined training pipeline	Advanced workflows

Starting a Training Job


from isa_sdk import ModelService
 
model_svc = ModelService("http://localhost:8082")
model_svc.set_auth_token(token)
 
job = model_svc.start_training({
    "algorithm_name": "apo",
    "model_id": "llama-3-8b",
    "dataset_id": "ds-001",
    "hyperparameters": {
        "learning_rate": 2e-5,
        "epochs": 3,
        "batch_size": 8,
        "lora_rank": 16,
    }
})
print(f"Training job started: {job['id']}")

LoRA Configuration

LoRA (Low-Rank Adaptation) reduces training cost by updating only a small subset of model parameters:

Parameter	Default	Description
`lora_rank`	16	Rank of low-rank matrices
`lora_alpha`	32	Scaling factor
`lora_dropout`	0.05	Dropout rate
`target_modules`	auto	Modules to apply LoRA

Monitoring Jobs


# Check job status
status = model_svc.get_training_job(job_id)
print(f"Status: {status['status']}, Progress: {status['progress']}%")
 
# List all jobs
jobs = model_svc.list_training_jobs()

Checkpoints

Training creates checkpoints at configurable intervals. Use the Console training page to view checkpoints and select the best one for deployment.