Python ML Model Training Evaluation

I am building a Windows-based computer-vision system that pulls images and related text from PDF and XML sources, then trains a deep-learning model to recognise those images with high accuracy. Everything happens in Python, so the entire training and evaluation pipeline—including data loaders, augmentation, model definition, loss functions, metrics, and experiment tracking—must be written in clean, modular Python code that runs smoothly on a local workstation or GPU server.

You will connect preprocessing routines to the raw documents, structure datasets for efficient loading, and iterate through model architectures (CNNs, transformers or other modern backbones). Precise reporting of training curves, confusion matrices, and F1/accuracy scores is essential; the project’s success hinges on demonstrably solid evaluation rather than a simple proof-of-concept run.

LLM-powered annotation suggestions are already available, so feel free to integrate them if it speeds up labelling or improves validation. Experience with PyTorch or TensorFlow, scikit-learn, OpenCV, and common tracking tools such as TensorBoard or Weights & Biases will be valuable.

Deliverables
• Reproducible Python codebase (data prep, training, evaluation)
• Trained weights and configuration files ready for inference
• Comprehensive evaluation report (metrics, plots, brief discussion)
• Setup guide so the entire workflow runs on another Windows machine without surprises

Acceptance criteria
– Model reaches the agreed-upon accuracy/F1 on the held-out test set
– Code executes with a single command after environment setup
– All outputs match the evaluation figures in the report