Document Type Classifier

Trains a classifier to identify the type of an academic document from its text content. The trained model is used by the Orchestrator API to select the correct type-specific prompt.

Mirrors fine_tune_subject — both share the same training strategies (utils/ml_strategies/); this module only supplies the type-specific dataset pipeline and model directories.

Run with:

./run_modules.sh fine_tune_type

Classification Target

Type	Spanish
Thesis	Tesis
Book	Libro
Article	Articulo
Conference Object	Objeto de conferencia

Module Structure

File	Responsibility
main.py	Backward-compatible entry point: runs make_dataset then train then test (interactive)
make_dataset.py	Data pipeline: create types CSV, download balanced PDFs (target 500/type), extract plain text (no tags)
create_types_csv.py	Builds the id→type CSV from SEDICI metadata
download_balance_pdfs.py	Downloads PDFs balanced across types
convert_pdfs_to_txt.py	Extracts plain text from downloaded PDFs (untagged, unlike the subject module which also cleans tags)
train.py	Training entry point — interactive menu or CLI (svm, xgboost, random_forest, embeddings, embeddings_knn, neural, minilm, all)
model_comparison_framework.py	Thin type-specific wrapper around the shared comparison framework
test.py	Test trained model(s) against a single PDF file

Dataset & Training

python -m fine_tune_type.make_dataset --all     # build dataset
python -m fine_tune_type.train all --compare    # train all strategies + comparison
python -m fine_tune_type.test /path/to/file.pdf # test on a single PDF

Data loading uses utils.ml_strategies.data_loader.load_csv_labels / create_dataset against CSV_FOLDER / CSV_TYPES and TXT_NO_TAGS_FOLDER, capping each label at SAMPLES_PER_TYPE (from constants.py, default 500) with min_frequency=5.

Each strategy is instantiated with an explicit model_dir=TYPE_MODEL_FOLDERS[...] (from constants.py) so it saves to a type-specific folder instead of the subject module's default — this is what lets the same SVMTrainingStrategy/XGBoostTrainingStrategy/etc. classes in utils/ml_strategies/strategies/ serve both classifiers. See Subject Classifier for the full strategy table.

Deployed Model

The TF-IDF + sklearn classifier (modelo_tipo_documento.pkl + vectorizador_tfidf.pkl) is the one loaded by the Orchestrator at runtime — see Orchestrator: Models Required.