New Release: Ultralytics v8.3.183

Discussion
, ,
1

Title: Ultralytics v8.3.183 — Train YOLO11 from NDJSON in One Command, YOLOE Docs Fixes, and More Reliable GPU CI

Summary
Train YOLO11 models directly from NDJSON datasets with a single command — automatic conversion, fast image downloading, improved YOLOE docs/examples, and more reliable GPU CI on GitHub.

What’s New in v8.3.183

  • NDJSON training support with automatic conversion and async downloads
  • YOLOE docs/examples improvements for clearer, prompt-free validation
  • GPU CI enabled on GitHub with reliability fixes for CUDA/ONNX/TensorRT

New Features

  • NDJSON training support (priority)
    • Built-in NDJSON→YOLO11 converter: convert_ndjson_to_yolo(ndjson_path, output_path=None) by @glenn-jocher
    • Trainer integration: pass a .ndjson file to model.train(data=…) and it auto-converts and generates data.yaml
    • Fast, resumable downloads: fetches images from URLs with async parallel workers and progress bar
    • Rich annotation support: handles detection boxes, segments, pose, OBB, and classification; writes the first available type found
    • Docs updated with NDJSON format examples and advantages
  • Minimal examples
    • Python
      • from ultralytics import YOLO
        model = YOLO(“yolo11n.pt”)
        model.train(data=“path/to/dataset.ndjson”, epochs=100, imgsz=640)
    • CLI
      • yolo detect train data=path/to/dataset.ndjson model=yolo11n.pt epochs=100 imgsz=640

Improvements

Bug Fixes

  • YOLOE docs example now uses correct weights and single-class validation, reducing confusion and validation noise
  • Stabilized CI pipelines for CUDA/ONNX/TensorRT paths to prevent flaky failures

Purpose & Impact

  • Simpler data onboarding: Train directly from NDJSON (with remote URLs) — no manual conversion required
  • Faster, scalable pipelines: Async, parallel downloads and streaming-friendly NDJSON boost large-scale dataset throughput
  • Broader task coverage: Unified support for detection, segmentation, pose, OBB, and classification
  • Fewer mistakes: Updated YOLOE examples ensure correct weights and cleaner validation outputs
  • More reliable CI: GPU tests run on GitHub with better dependency handling, improving confidence in CUDA/ONNX/TensorRT paths

PRs in This Release

New Contributors

Get Started

  • pip install -U ultralytics
  • Try NDJSON training:
    • yolo detect train data=path/to/dataset.ndjson model=yolo11n.pt epochs=100 imgsz=640

Links

Call for Feedback
Give v8.3.183 a spin, especially the NDJSON training flow and updated YOLOE examples. Share your experience, datasets, and any edge cases you encounter — your feedback helps the whole YOLO community and the Ultralytics team keep improving.