Skip to main content
Run SAM or YOLO inference across all unlabeled items in a project to generate draft annotations. Review and refine the results instead of labeling from scratch.

​
Overview

Batch auto-labeling takes a project, runs a model against every unlabeled item, and creates draft annotations that annotators can accept, edit, or discard. This dramatically reduces manual labeling time for projects where a pre-trained model can provide a reasonable starting point. 
Start batch job
      │
      ▼
Fetch unlabeled items
      │
      ▼
Run inference per item ──► Below threshold? → Skip
      │
      ▼
Create draft result
      │
      ▼
Emit webhook event

​
Quickstart

import httpx
import time

# Start a batch auto-label job
resp = httpx.post(
    "https://api.avala.ai/api/v1/projects/proj_abc123/auto-label/",
    headers={"X-Avala-Api-Key": "avk_..."},
    json={
        "model_type": "yolo",
        "confidence_threshold": 0.6,
        "dry_run": False,
    },
)
job = resp.json()
print(f"Job started: {job['uid']}")

# Poll for progress
while job["status"] in ("pending", "running"):
    time.sleep(5)
    resp = httpx.get(
        f"https://api.avala.ai/api/v1/auto-label-jobs/{job['uid']}/",
        headers={"X-Avala-Api-Key": "avk_..."},
    )
    job = resp.json()
    print(f"Progress: {job['progress_pct']}% ({job['processed_items']}/{job['total_items']})")

print(f"Done! {job['successful_items']} items labeled, {job['skipped_items']} skipped")

​
API Reference

​
Create Job

POST /api/v1/projects/{project_uid}/auto-label/
ParameterTypeDefaultDescription
model_typestring"yolo"Inference model: "sam3" or "yolo".
confidence_thresholdfloat0.5Minimum confidence to accept a prediction (0.0-1.0).
labelsstring[][]Filter to specific labels. Empty = all labels.
dry_runbooleanfalseRun inference without creating results.
Response: 202 Accepted 
{
  "uid": "a1b2c3d4-...",
  "status": "pending",
  "model_type": "yolo",
  "confidence_threshold": 0.6,
  "total_items": 0,
  "processed_items": 0,
  "successful_items": 0,
  "failed_items": 0,
  "skipped_items": 0,
  "progress_pct": 0.0,
  "created_at": "2026-02-23T10:30:00Z"
}

​
Get Job Status

GET /api/v1/auto-label-jobs/{uid}/
Returns the current state of the job including progress counters.

​
List Jobs

GET /api/v1/auto-label-jobs/?project={project_uid}&status={status}
ParameterTypeDescription
projectstring?Filter by project UID.
statusstring?Filter by status: pending, running, completed, failed, cancelled.

​
Cancel Job

DELETE /api/v1/auto-label-jobs/{uid}/
Cancels a running or pending job. Items already processed are kept.

​
Models

​
YOLO (Object Detection)

Best for: Detecting and labeling objects with bounding boxes.
  • Generates bounding box annotations
  • Works well for common object categories
  • Fast inference (~50ms per image)

​
SAM (Segmentation)

Best for: Precise object boundaries and segmentation masks.
  • Generates segmentation annotations
  • Better boundary accuracy than bounding boxes
  • Slower inference (~200ms per image)

​
Configuration Guide

​
Confidence Threshold

The confidence_threshold controls how many predictions are accepted:
ThresholdEffect
0.3More predictions, more noise. Good for initial exploration.
0.5Balanced. Recommended starting point.
0.7Fewer predictions, higher quality. Good for production.
0.9Only very confident predictions. Minimal false positives.

​
Dry Run

Use dry_run: true to preview what the model would label without creating any results: 
curl -X POST .../projects/proj_abc123/auto-label/ \
  -H "X-Avala-Api-Key: avk_..." \
  -d '{"model_type": "yolo", "confidence_threshold": 0.5, "dry_run": true}'
Check the successful_items count to see how many items would be labeled.

​
Webhook Events

Auto-label jobs emit webhook events on completion:
EventTrigger
auto_label.completedJob finished successfully
auto_label.failedJob failed with an error
{
  "event": "auto_label.completed",
  "payload": {
    "job_uid": "a1b2c3d4-...",
    "project_uid": "proj_abc123",
    "status": "completed",
    "total_items": 500,
    "successful": 423,
    "failed": 2,
    "skipped": 75,
    "dry_run": false
  }
}

​
Limitations

  • Only one auto-label job can run per project at a time
  • Maximum 5,000 items per job
  • Requires a running SageMaker inference endpoint
  • Draft results are attributed to the Avala bot user
  • Only items without existing results are processed

​
Next Steps