yangwenli661/PMD-model-artifacts

PMD Model Artifacts

Private university-submission repository for Political Misinformation Detection in Somali Text Using Machine Learning.

Source code

The corresponding Flask application is maintained at:

• https://github.com/Al-Abaaska/webapp

Runtime classifiers

The web application loads eight classifiers:

1. Logistic Regression
2. Support Vector Machine (SVM)
3. Random Forest
4. Multinomial Naive Bayes
5. BiLSTM
6. CNN-BiLSTM
7. SomBERTa (runtime directory: somerta_model)
8. SomBERTb (runtime directory: sombertb_model)

The repository also contains the fitted TF-IDF vectorizer and shared Keras tokenizer required for inference.

Label mapping

• 0: Reliable
• 1: Misinformation
• 2: Misleading

Installation

Authenticate to Hugging Face, then download this repository directly into the web application's ml_models directory:

hf download yangwenli661/PMD-model-artifacts --local-dir ml_models

Python alternative:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="yangwenli661/PMD-model-artifacts",
    local_dir="ml_models",
)

For this private repository, the downloader must be authenticated or provide a read token through a secure environment variable. Never commit the token.

Integrity verification

checksums.sha256 records the SHA-256 digest of every uploaded runtime artifact. model_manifest.json records file sizes, paths, label mapping, and source-code location.

Scope and limitations

These files reproduce the models used by the PMD prototype. An AI label and confidence score are not a verified fact-checker decision. Model performance and limitations should be interpreted using the accompanying thesis and evaluation notebook.