Boost YOLOv26 mAP Beyond 85%
Budget / Salary$10–11
TypeFreelance project
LocationRemote
Posted1 hour ago
I’m looking for a hands-on refinement of my existing Kaggle notebook that trains two YOLOv26 models. Validation mAP currently stalls just under 85 %, and I need that figure pushed past the 85 % mark under the standard evaluation settings. The present training/validation split must stay exactly as it is, and every product type carries equal weight, so improvements can’t come at the expense of any single class. Highest overall mAP is the only success metric that counts.
Feel free to experiment with hyper-parameter sweeps, smarter augmentations, anchor re-calculation, label-smoothing tricks, or even a lightweight ensemble—as long as the full pipeline still runs end-to-end inside a clean Kaggle runtime and remains easy to reproduce.
Deliverables
• A revised Kaggle notebook running start-to-finish that reaches ≥ 85 % mAP, with concise comments explaining each change
• Final trained .pt weights for both models saved to the notebook’s output
• Clear evaluation visuals generated in-notebook: loss curves, precision-recall plots, confusion matrix, and the mAP summary that proves the target
• A small standalone Python script (OpenCV acceptable) that loads the exported weights and performs real-time webcam inference on the same class set
Acceptance criteria
• Notebook runs without manual intervention on a fresh Kaggle GPU session
• Reported mAP ≥ 85 % on the maintained validation split
• Script delivers real-time predictions at usable FPS on a standard laptop webcam
If you can deliver repeatable gains and tidy, well-commented code, I’m ready to get started right away.
Feel free to experiment with hyper-parameter sweeps, smarter augmentations, anchor re-calculation, label-smoothing tricks, or even a lightweight ensemble—as long as the full pipeline still runs end-to-end inside a clean Kaggle runtime and remains easy to reproduce.
Deliverables
• A revised Kaggle notebook running start-to-finish that reaches ≥ 85 % mAP, with concise comments explaining each change
• Final trained .pt weights for both models saved to the notebook’s output
• Clear evaluation visuals generated in-notebook: loss curves, precision-recall plots, confusion matrix, and the mAP summary that proves the target
• A small standalone Python script (OpenCV acceptable) that loads the exported weights and performs real-time webcam inference on the same class set
Acceptance criteria
• Notebook runs without manual intervention on a fresh Kaggle GPU session
• Reported mAP ≥ 85 % on the maintained validation split
• Script delivers real-time predictions at usable FPS on a standard laptop webcam
If you can deliver repeatable gains and tidy, well-commented code, I’m ready to get started right away.
Apply on Freelancer →
Project sourced from Freelancer.com. Applications happen directly on the original platform — we never collect your data.