Training

Training allows organizations to fine-tune AI models using their accumulated context (embeddings).

Concepts

Concept	Description
Training Job	A model fine-tuning task
LoRA/DoRA	Parameter-efficient fine-tuning methods
Adapter	The trained model weights
Rollback	Revert to previous adapter

Training Lifecycle

Pending → Running → Completed/Failed

Status	Description
pending	Job queued
running	Training in progress
completed	Successfully finished
failed	Training failed
cancelled	User cancelled

Types

From types/trainings.ts:

interface StartTrainingRequest {
  name: string;
  description?: string;
  epochs?: number;
  learningRate?: number;
  batchSize?: number;
  // Additional config options
}

interface StartTrainingResponse {
  jobId: string;
  status: TrainingStatus;
}

interface ListJobsResponse {
  jobs: TrainingJob[];
}

interface TrainingJob {
  id: string;
  name: string;
  description?: string;
  status: TrainingStatus;
  progress?: number;
  metrics?: TrainingMetrics;
  createdAt: string;
  startedAt?: string;
  completedAt?: string;
}

interface TrainingMetrics {
  loss?: number;
  accuracy?: number;
  epoch?: number;
}

type TrainingStatus = 
  | "pending"
  | "running" 
  | "completed"
  | "failed"
  | "cancelled";

Hooks

useTrainings

hooks/trainings/use-trainings.ts

List jobs and create new training.

const {
  jobs,            // TrainingJob[]
  isLoading,
  createJob        // (data: StartTrainingRequest) => Promise<StartTrainingResponse>
} = useTrainings(organizationSlug);

API Calls: - GET /train/jobs - List all jobs - POST /train - Start new training job

useTraining

hooks/trainings/use-training.ts

Manage a single training job.

const {
  job,             // TrainingJob
  status,          // TrainingStatus (polled)
  isLoading,
  isPolling,       // Still running
  cancelJob,       // () => Promise<void>
  rollbackJob      // () => Promise<void>
} = useTraining(organizationSlug, jobId);

API Calls: - GET /train/jobs/{id} - Job details - GET /train/jobs/{id}/status - Job status (polled) - DELETE /train/jobs/{id} - Cancel/delete job - POST /train/rollback/{id} - Rollback to previous

Polling:

// Polls status while job is running
refetchInterval: (job) => 
  ["pending", "running"].includes(job?.status) ? 5000 : false

Pages

Training Overview

/portal/[slug]/[teamID]/training

Tutorial cards for training workflow: 1. Build context (embeddings) 2. Configure training job 3. Run training 4. Deploy adapter

New Training Job

/portal/[slug]/[teamID]/training/new

Configuration form: - Job name - Description - Training parameters: - Epochs - Learning rate - Batch size

const handleSubmit = async (data: StartTrainingRequest) => {
  const response = await createJob(data);
  router.push(`/portal/${slug}/${teamID}/training/jobs/${response.jobId}`);
};

Jobs List

/portal/[slug]/[teamID]/training/jobs

Table of all training jobs: - Name, status, progress - Created/completed dates - Actions (view, cancel)

Job Detail

/portal/[slug]/[teamID]/training/jobs/[jobID]

Job monitoring page: - Status indicator - Progress bar (while running) - Metrics (loss, accuracy per epoch) - Rollback action (if completed) - Cancel action (if running) - Delete action

export default function JobDetailPage({ params }) {
  const { job, status, rollbackJob, cancelJob } = useTraining(
    params.slug,
    params.jobID
  );

  return (
    <div>
      <JobHeader job={job} status={status} />

      {status === "running" && (
        <ProgressBar value={job.progress} />
      )}

      {job.metrics && (
        <MetricsDisplay metrics={job.metrics} />
      )}

      <JobActions
        status={status}
        onCancel={cancelJob}
        onRollback={rollbackJob}
      />
    </div>
  );
}

Components

JobsTable

components/tables/trainings/jobs-table.tsx

Data table for training jobs: - Sortable columns - Status badges - Progress indicators - Action buttons

Access Control

Training is restricted to admin+ roles:

// In PortalSidebar
{isAdmin && (
  <SidebarMenuItem>
    <Link href={`/portal/${slug}/${teamID}/training`}>
      Training
    </Link>
  </SidebarMenuItem>
)}

// In training pages
<RoleProvider requireRole="admin">
  <TrainingPage />
</RoleProvider>

Workflow Example

1. Prepare Context
2. Upload relevant files to embeddings
3. Embed session timelines

4. Ensure sufficient training data

5. Configure Job

6. Go to /training/new
7. Name the job (e.g., "Q1 Game Analysis Model")

8. Optionally adjust hyperparameters

9. Start Training

10. Submit configuration
11. Job enters pending state

12. Redirected to job detail page

13. Monitor Progress

14. Status updates via polling
15. View training metrics

16. Wait for completion

17. Handle Results

18. Success: Adapter deployed automatically

19. Failed: Review logs, adjust config, retry

20. Rollback (if needed)

21. If new adapter underperforms
22. Click rollback to restore previous

Training Parameters

Parameter	Description	Default
epochs	Training iterations	Varies
learningRate	Step size for optimization	Varies
batchSize	Samples per gradient update	Varies

Metrics

During and after training:

Metric	Description
loss	Training loss value
accuracy	Model accuracy
epoch	Current epoch number
progress	Percentage complete