jmptorralba/gabai-models

GABai models — AI Reading Tutor (capstone deployment)

Trained model artifacts for GABai, an offline-capable AI reading tutor for Filipino learners aged 5–7. Code, docs, and setup instructions: https://github.com/jmptorralba/gabai_local

From the capstone project "Bridging Educational Gaps: A Scalable AI Model for Predicting Reading Proficiency and Delivering Adaptive Learning Experiences" (J.M.P. Torralba, University of Asia and the Pacific, MABA, 2025).

Contents

File	What it is
rf_pipeline.joblib	Random Forest reading-proficiency classifier (scikit-learn 1.6.1 — keep that version), trained on PISA Philippines 2018/2022 socio-demographic data. Predicts Critical / Not Critical reading proficiency from 9 background questions. ~66% accurate — the app treats it as a starting point with a mandatory human override.
flan_t5_base_gabai_mcqa/	LoRA adapter (r=16, q+v) for google/flan-t5-base, fine-tuned on FairytaleQA to generate a comprehension question + correct answer from a children's story.

Not included here: the story-generation model. GABai uses the Q4_K_M GGUF of Meta-Llama-3.1-8B-Instruct, downloaded from QuantFactory/Meta-Llama-3.1-8B-Instruct-GGUF (subject to the Llama 3.1 Community License). The GitHub repo's download_models.py fetches everything, including this repo, and verifies checksums.

Usage

Don't load these directly — they are wired into the GABai pipeline, which adds the exact input encoding the classifier requires and a governance layer (safety/readability/structure gates) around all generated content. Clone the GitHub repo and run download_models.py.

License & data notes

• The QA adapter was fine-tuned on FairytaleQA (CC BY-NC 4.0) — this repo is therefore non-commercial (research and education use).
• The classifier was trained on publicly available OECD PISA data; it uses only anonymous background categories, no personal data.
• Base model google/flan-t5-base is Apache 2.0 (downloaded separately at first run).