How abdulmohsena/faseeh works
Main Components
• Training Module – Represented by files in the "train" folder (main.ipynb, cloud_train.ipynb, and train_modal.py). This module is responsible for data ingestion (even if not explicitly in the file tree, the training scripts would include preprocessing), training the translation model, and saving model checkpoints.
• Evaluation & Inference Module – Captured in the "eval" folder (metrics.ipynb) and via Inference_Strategy.ipynb. This part evaluates the trained model using different metrics (BLEU, COMET, BERTSCORE, etc.) and demonstrates how inference is carried out.
• Model Repository/Packaging – The "repo" directory (with its contained Faseeh module/directory) likely houses the model’s code or a packaged version, which might be used for deployment or integration (as also indicated by the hug link to the model repository on Hugging Face).
• External Resources – Although not explicitly detailed in the file tree, the README mentions a connection to a Hugging Face model repository. Also, the "requirements.txt" file signals dependency management for various libraries/frameworks used in the project.
Out the Relationships and Data Flow
• “Trained Model” output to the Evaluation & Inference Module, indicating that the model undergoes performance assessment.
• Optionally, indicate external dependencies like Python libraries (from requirements.txt) and external services (the Hugging Face model repository).
Architectural Patterns or Principles
• Separation of Concerns – There is a clear division between training, evaluation, and deployment/inference.
• Modularity – Different notebooks and scripts are grouped by their purpose (training in one folder, evaluation in another, and packaging/integration in the “repo”). This modular approach should be reflected by grouping those components together.