GroovyMorning_RLAC

Experiment Tracking and Monitoring with Weights & Biases (WandB)

Training artificial intelligence models is an iterative process requiring detailed metric analysis. To ensure transparency and reproducibility of results, GroovyMorning uses Weights & Biases (WandB) as a centralized monitoring platform.

Why WandB?

As part of the R&D on chronicle detection, we test dozens of configurations (hyperparameters, model architectures, dataset sizes). WandB allows us to:

  1. Visualize in Real-Time: Track the evolution of Loss, Precision, Recall, and F1-Score during training.
  2. Compare Runs: Overlay curves from different experiments to instantly identify which model converges best.
  3. Debug Training: Quickly identify issues with overfitting or unstable gradients.
  4. Organize by Tags: Use tags (e.g., wavlm, cnn, jingle) to filter experiments by architecture type.

w-and-b.png

Technical Integration

The monitoring is directly integrated into the training scripts of the 4.rd/1.IAChronicleDetector module.

Initialization

Each training session begins with an initialization linked to the global project RLAC or IAChronicleDetection:

import wandb
wandb.init(
    project="IAChronicleDetection",
    config={
        "learning_rate": 2e-5,
        "architecture": "CamemBERT",
        "dataset": "RadioFrance-V2"
    }
)

Automatic Reporting

Thanks to integration with the transformers and accelerate libraries, metrics are sent automatically at each step (logging_steps). The report_to="wandb" parameter in TrainingArguments is enough to activate this link.

Dashboards

The WandB interface serves as a “control tower” for the project’s AI engineers. It allows generating performance reports that serve as the basis for selecting the final model to be deployed in production on Hugging Face.

This rigor in tracking ensures that every improvement made to the GroovyMorning detection engine is scientifically validated and documented.

Version française : Suivi d’Expériences et Monitoring avec Weights & Biases (WandB)

This site is open source. Improve this page.