YOLO World has extremely low training metrics on LVIS.

[hongliangchen211-source]

YOLO World has extremely low training metrics on LVIS.

[UltralyticsAssistant]

👋 Hello @hongliangchen211-source, thank you for your interest in Ultralytics 🚀! We recommend a visit to the Docs for new users, where you can find many Python and CLI usage examples and where many common questions may already be answered.

This is an automated response 🤖, and an Ultralytics engineer will also assist you soon.

For this report about extremely low YOLO-World training metrics on LVIS 🐛, please provide a minimum reproducible example, including:

• The exact command or code used for training and validation
• Your ultralytics version, Python version, PyTorch version, and hardware details
• The model configuration or checkpoint used
• The LVIS dataset configuration and a description of any dataset modifications
• Complete training logs and validation output
• Representative dataset images with labels, if possible

Please also provide as much information as possible about your custom training setup and verify that you are following our Tips for Best Training Results.

Join the Ultralytics community where it suits you best. For real-time chat, head to Discord 🎧. Prefer in-depth discussions? Check out Discourse. Or dive into threads on our Subreddit to share knowledge with the community.

Upgrade

Upgrade to the latest ultralytics package including all requirements in a Python>=3.8 environment with PyTorch>=1.8 to verify your issue is not already resolved in the latest version:

pip install -U ultralytics

Environments

YOLO may be run in any of the following up-to-date verified environments (with all dependencies including CUDA/CUDNN, Python and PyTorch preinstalled):

• Notebooks with free GPU:
• Google Cloud Deep Learning VM. See GCP Quickstart Guide
• Amazon Deep Learning AMI. See AWS Quickstart Guide
• Docker Image. See Docker Quickstart Guide

Status

If this badge is green, all Ultralytics CI tests are currently passing. CI tests verify correct operation of all YOLO Modes and Tasks on macOS, Windows, and Ubuntu every 24 hours and on every commit.

[Y-T-G]

You should use the code that's provided in docs to reproduce:

https://docs.ultralytics.com/models/yolo-world#launch-training-from-scratch

[hongliangchen211-source]

However, the download link for the annotated file seems to be broken.

[Y-T-G]

Can you post which link is broken? Maybe your internet blocks it

[hongliangchen211-source]

I would also like to ask how to reproduce the COCO fine-tuning experiment described in the paper.

[glenn-jocher]

For COCO fine-tuning, load the pretrained checkpoint and use the standard WorldTrainer, e.g. YOLOWorld("yolov8s-worldv2.pt").train(data="coco.yaml", epochs=100, imgsz=640); to reproduce the paper’s exact result, please specify the table/model variant since its batch size, schedule, and other hyperparameters must match that experiment.

[rishig47-dev]

If the YOLO-World training metrics on LVIS are extremely low, I would first verify the evaluation pipeline before changing the model or training recipe.

LVIS has a much larger vocabulary than COCO (1,203 categories), so the class-name/text-embedding setup is important for YOLO-World. The current Ultralytics implementation explicitly handles LVIS class names during validation because YOLO-World normally starts with COCO-style classes.

I would check these points first:

1. Use the LVIS validation/minival split correctly. For YOLO-World training, Ultralytics' WorldTrainer specifically switches to the LVIS minival split for final evaluation.

2. Make sure the model is using the LVIS class names. Don't leave the model configured with the default COCO 80 classes when evaluating LVIS. The validator needs to generate text embeddings for the LVIS class names.

3. Check the Ultralytics version. There have been LVIS/YOLO-World validation issues involving class indexing and zero metrics, so reproducing the result with a current Ultralytics version is important before modifying detect/val.py manually.

4. Compare against the pretrained checkpoint first.** Run validation on the pretrained YOLO-World checkpoint before training. If the pretrained model also produces near-zero LVIS metrics, the problem is much more likely to be the validation/class mapping pipeline than your training recipe.

5. Don't immediately modify DetectionValidator.init_metrics(). If changing the validator makes the code run but produces all-zero metrics, that is a strong indication that the underlying class mapping/evaluation setup is still incorrect rather than the model simply having poor accuracy.

I would first reproduce the official YOLO-World + LVIS workflow using the current Ultralytics implementation, then compare pretrained vs. trained results. Once the validation pipeline is confirmed, you can investigate learning rate, training data, augmentation, and checkpoint quality.

For reference, Ultralytics' current YOLO-World documentation uses LVIS minival as the validation dataset when reproducing YOLO-World results.

[pderrenger]

Please share the exact training/validation command, ultralytics version, pretrained checkpoint, resulting metrics, and full text logs so we can reproduce this. Current YOLO-World validation sets embeddings from the dataset class names, while WorldTrainerFromScratch evaluates LVIS on minival; see the YOLO-World validation reference.