EphraimShikanga/agri-mtl-research

MTL Transformer research pipeline for Kenya crop recommendation system # Research — Kenya Agricultural MTL Transformer Research code for the Multi-Task Learning Transformer that powers the crop recommendation system. This folder contains everything needed to reproduce the data pipeline, model training, and analysis — separate from the production Django server (`neuro_symbolic_engine/`). ## Directory Structure ``` research/ ├── scripts/ │ ├── crop_metadata.py ← Shared constants (22 crops, harvest months, etc.) │ ├── data_acquisition/ │ │ └── download_weather_data.py ← NASA POWER weather data downloader │ ├── preprocessing/ │ │ ├── preprocess_food_prices.py ← Clean WFP food prices │ │ ├── preprocess_yield.py ← Filter FAO yield data to Kenya │ │ ├── preprocess_ag_production.py← Agricultural production preprocessing │ │ ├── add_confidence_weights.py ← Add confidence weights to yield data │ │ ├── build_weather_sequences.py ← 12-month sliding weather windows │ │ ├── build_price_sequences.py ← 12-month price windows + 6-month look-ahead │ │ ├── build_yield_outputs.py ← Yield baseline computation │ │ ├── build_training_forms.py ← Merge all datasets into training samples │ │ └── compute_reference_prices.py← National avg price sequences for inference │ └── model/ │ └── mtl_transformer.py ← Full model: architecture + training loop ├── notebooks/ │ ├── 01_data_exploration.ipynb ← EDA on raw datasets │ ├── 02_preprocessing.ipynb ← Interactive preprocessing walkthrough │ ├── 03_train_val_test_split.ipynb ← Train/val/test split │ └── 04_mtl_transformer_training.ipynb ← Model training (run on Colab GPU) ├── data/ │ ├── raw/ ← Original source datasets │ │ ├── Crop_recommendation.csv │ │ ├── yield_df.csv │ │ ├── wfp_food_prices_ken.csv │ │ ├── wfp_markets_ken.csv │ │ ├── kenya_weather_all_counties.csv │ │ ├── Kenyas_Agricultural_Production.xlsx │ │ └── weather_data/ …