GCC Education AI: 10-Model Suite
This repository contains ten small scikit-learn models for education-sector questions in the Gulf Cooperation Council (GCC) region, with a Kuwait focus. They cover:
- student performance
- university dropout
- school enrollment demand
- teacher retention
- Arabic language proficiency
- STEM readiness
- special-needs support demand
- private tutoring demand
- scholarship eligibility
- EdTech adoption
A simple AutoML loop picked each model by comparing five scikit-learn algorithms. All training data was generated synthetically by the training script (see Training data). The suite is a prototype and teaching reference for education analysts, researchers and developers. It is not a validated tool for decisions about individual students or teachers.
Author: AgenThink, Kuwait City
Models
"Selected algorithm" means the candidate with the best 3-fold cross-validation score on the training split. Test metrics come from a 20% hold-out split. All numbers are copied from automl_results.json.
| # | Model file | Task | Type | Selected algorithm | CV score | Test score | Test MAE |
|---|---|---|---|---|---|---|---|
| 1 | model_student_performance.pkl |
Student Academic Performance (GPA, 30–100) | Regression | Ridge | R² 0.8486 | R² 0.8524 | 3.2315 |
| 2 | model_dropout_risk.pkl |
University Dropout Risk | Classification (2 classes) | GradientBoosting | Acc 0.9692 | Acc 0.9700 | – |
| 3 | model_enrollment_demand.pkl |
School Enrollment Demand (students) | Regression | Ridge | R² 0.9328 | R² 0.9303 | 80.6795 |
| 4 | model_teacher_retention.pkl |
Teacher Retention (leaves / stays) | Classification (2 classes) | GradientBoosting | Acc 0.9526 | Acc 0.9592 | – |
| 5 | model_arabic_proficiency.pkl (see note) |
Arabic Language Proficiency (20–100) | Regression | GradientBoosting | R² 0.6984 | R² 0.7126 | 4.4354 |
| 6 | model_stem_readiness.pkl |
STEM Career Readiness | Classification (2 classes) | LogisticRegression | Acc 0.8640 | Acc 0.8688 | – |
| 7 | model_special_needs.pkl (see note) |
Special Needs Support Demand (5–800) | Regression | Ridge (per JSON) | R² 0.2853 | R² 0.2568 | 15.0056 |
| 8 | model_tutoring_demand.pkl |
Private Tutoring Demand Index (5–150) | Regression | GradientBoosting | R² 0.8337 | R² 0.8370 | 7.9619 |
| 9 | model_scholarship.pkl |
Scholarship Eligibility | Classification (2 classes) | GradientBoosting | Acc 0.9499 | Acc 0.9425 | – |
| 10 | model_edtech_adoption.pkl |
EdTech Adoption | Classification (2 classes) | GradientBoosting | Acc 0.9445 | Acc 0.9458 | – |
Note: Model/metadata mismatch (models 5 and 7). The uploaded pickles for these two models do not match the training script or automl_results.json:
| File | Pickled estimator | Hyperparameters | Inputs | Mismatch with JSON |
|---|---|---|---|---|
model_arabic_proficiency.pkl |
GradientBoostingRegressor |
n_estimators=100, max_depth=5 |
10 unnamed columns (n_features_in_=10, no feature_names_in_) |
JSON describes an 80-tree, depth-4 model over 20 named features |
model_special_needs.pkl |
GradientBoostingRegressor |
n_estimators=100, max_depth=5 |
8 unnamed columns (n_features_in_=8, no feature_names_in_) |
JSON reports a Ridge model over 20 named features |
These two files were evidently produced by a separate run that is not included in this repository. Their input columns, column order and accuracy are unknown, and the metrics in the table do not describe them. Treat them as unusable until they are regenerated with gcc_education_ai.py. The other eight pickles match the script: same estimator class, same hyperparameters, and the same 18–20 named features in the same order.
Candidate algorithms, with hyperparameters fixed in the script:
- Regression:
RandomForestRegressor(n_estimators=80, max_depth=10),GradientBoostingRegressor(n_estimators=80, max_depth=4, learning_rate=0.1),ExtraTreesRegressor(n_estimators=80, max_depth=10),Ridge(alpha=10.0),DecisionTreeRegressor(max_depth=8). - Classification: the matching classifiers, plus
LogisticRegression(max_iter=300).
All candidates use random_state=42. automl_results.json records every candidate's CV score under all_scores. The pickles were saved with scikit-learn 1.8.0.
Classification labels, all binary with 1 as the positive class. The positive share follows from the quantile threshold used to build each label:
- Dropout Risk: 1 = drops out (about 18% positive)
- Teacher Retention: 1 = teacher leaves (about 22% positive)
- STEM Readiness: 1 = STEM-ready (about 35% positive)
- Scholarship: 1 = awarded (about 35% positive)
- EdTech Adoption: 1 = adopts (about 45% positive)
Dashboard
The training script generates this dashboard. It shows per-model scores, selected feature importances, views of the synthetic data and a heatmap comparing the candidate algorithms.
Repository files
| File | Size | Description |
|---|---|---|
gcc_education_ai.py |
44 KB | Full pipeline: synthetic data generation, AutoML selection, evaluation, dashboard, model export |
automl_results.json |
16 KB | Per-model task type, selected algorithm, CV/test metrics, all candidate scores, feature list, feature importances, short domain note |
gcc_education_ai_dashboard.png |
952 KB | Results dashboard |
model_student_performance.pkl |
1 KB | Ridge regressor |
model_dropout_risk.pkl |
203 KB | GradientBoostingClassifier |
model_enrollment_demand.pkl |
1 KB | Ridge regressor |
model_teacher_retention.pkl |
202 KB | GradientBoostingClassifier |
model_arabic_proficiency.pkl |
465 KB | GradientBoostingRegressor, 10 unnamed inputs. Does not match the script or JSON (see above) |
model_stem_readiness.pkl |
2 KB | LogisticRegression |
model_special_needs.pkl |
467 KB | GradientBoostingRegressor, 8 unnamed inputs. Does not match the script or JSON (see above) |
model_tutoring_demand.pkl |
205 KB | GradientBoostingRegressor |
model_scholarship.pkl |
204 KB | GradientBoostingClassifier |
model_edtech_adoption.pkl |
205 KB | GradientBoostingClassifier |
README.md |
– | This model card |
Training data
No real-world data was used. For each task, gcc_education_ai.py generates 12,000 synthetic rows from NumPy random distributions with seed 42. The distributions (uniform, normal, lognormal, beta, Poisson, binomial and categorical draws) use ranges chosen to look plausible for Kuwait and the GCC. For example, the tutoring rate is set to 55% and the Kuwaiti/expat shares are set by hand.
- Regression targets are hand-written formulas over a few input columns, plus Gaussian noise, then clipped. For example, GPA is a weighted sum of previous GPA, study hours, attendance, tutoring, parental education, teacher experience, screen time, health absences and a Ramadan effect.
- Classification labels come from thresholding a hand-written risk or eligibility score at a fixed quantile.
- Unused columns: several inputs do not appear in any generating formula. In Enrollment Demand,
covid_effectandtutor_substituteare generated but are not model inputs. - Train/test protocol: each task uses an 80/20 train/test split (
random_state=42). Model selection uses 3-fold CV on the 80% portion. The selected model is refit on that 80% and scored on the 20% hold-out.
The gcc_note strings in automl_results.json are contextual remarks by the author, for example "Private tutoring rate 55%+ in Kuwait". The code does not derive them from data, and this card does not verify them.
Features / inputs
Columns must match the names and order below. The eight consistent pickles store feature_names_in_. Categorical codes come from LabelEncoder, so they follow alphabetical order.
- Student Performance (19 features):
school_enc, grade_level, age, is_female, is_kuwaiti, class_size, teacher_experience, parental_education, family_income_tier, study_hours_daily, tutoring, screen_time_hours, attendance_pct, arabic_at_home, prev_gpa, extracurricular, health_absences, summer_school, digital_resourcesschool_enc: 0=International, 1=Private-Arabic, 2=Private-English, 3=Public-Arabic, 4=Public-English
- Dropout Risk (20 features):
uni_enc, faculty, year_of_study, entry_gpa, current_gpa, is_female, is_kuwaiti, scholarship, works_part_time, commute_hours, attendance_pct, failed_courses, family_pressure, financial_stress, social_integration, mental_health_score, language_barrier, wrong_major, advisor_meetings, oil_job_pulluni_enc: 0=ACK, 1=AOU, 2=AUK, 3=Abroad, 4=GUST, 5=Gulf University, 6=Kuwait University
- Enrollment Demand (18 features):
school_type_enc, district, year, month, population_growth, expat_inflow, oil_price_usd, birth_rate_lag6yr, private_school_fees, public_capacity, new_residential_dev, expat_policy_strict, summer_exodus, online_alt_pct, nationalization_push, gdp_per_capita, fertility_rate, teachers_available - Teacher Retention (20 features):
subject_enc, school_type_enc, years_teaching, age, is_kuwaiti, is_female, salary_kwd, class_load, students_per_class, admin_burden, student_behaviour, parental_pressure, professional_dev, management_support, visa_renewals_left, family_in_gcc, better_offer, housing_provided, commute_km, ramadan_workloadsubject_enc: 0=Arabic, 1=Arts, 2=English, 3=IT, 4=Islamic, 5=Math, 6=PE, 7=Science, 8=Social, 9=Special
- Arabic Proficiency: the script defines 20 features (
dialect_enc, age, grade_level, arabic_at_home_hrs, school_arabic_hrs, is_native_arabic, quran_study, arabic_media_hrs, english_dominance, reading_habits, teacher_quality, parental_arabic_edu, digital_arabic_use, private_arabic_tutor, msa_exposure, literature_exposure, bilingual_school, travel_arab_countries, social_media_arabic, family_size). The uploaded pickle expects 10 unnamed inputs, whose identity is unknown. - STEM Readiness (20 features):
math_score, science_score, it_score, english_score, grade_level, is_female, stem_interest, stem_extracurricular, coding_experience, robotics_club, parental_stem_bg, stem_teacher_quality, lab_access, national_competition, tutoring_stem, critical_thinking, problem_solving, oil_sector_aspiration, university_stem_intent, scholarship_available - Special Needs Demand: the script defines 20 features (
district, school_count, total_students_k, consanguinity_rate, diagnosis_autism, diagnosis_adhd, diagnosis_learning, diagnosis_physical, diagnosis_speech, specialist_count, resource_rooms, budget_kwd_k, awareness_score, parental_stigma, year, early_intervention, expat_student_pct, technology_support, govt_initiative, ngo_support). The uploaded pickle expects 8 unnamed inputs, whose identity is unknown. - Tutoring Demand (20 features):
district, grade_level, subject_enc, exam_season, ramadan_period, school_type_enc, family_income_tier, avg_class_performance, parent_education, both_parents_working, prev_year_results, online_tutoring_avail, tutor_hourly_rate, school_quality_score, cultural_pressure, num_siblings, oil_price_env, summer_program, university_prep, population_k - Scholarship (20 features):
schol_enc, gpa, is_kuwaiti, is_female, family_income_kwd, parent_govt_employee, target_country, intended_major, extracurricular_score, leadership_roles, volunteer_hours, language_test_score, recommendation_score, interview_score, disability_status, orphan_status, siblings_on_scholarship, prev_scholarship, stem_major, oil_sector_majorschol_enc: 0=Corporate, 1=Government, 2=International, 3=Merit, 4=Need-Based, 5=University
- EdTech Adoption (20 features):
school_type_enc, grade_level, teacher_age, teacher_tech_comfort, device_per_student, internet_quality, it_support_staff, admin_mandate, training_received, student_engagement, parent_digital_lit, platform_usability, content_arabic, post_covid, govt_initiative, budget_edtech_kwd, pilot_success, ai_tools_awareness, oil_price_env, region_connectivity
Some integer-coded inputs are random integers with no defined category mapping: school_type_enc in models 3, 4, 8 and 10, plus faculty, district and subject_enc in model 8. Ordinal scores are integers from 1 to 5, and flags are 0 or 1.
How to use
The script saved the models with joblib.dump. Load them with joblib, using scikit-learn 1.8.0.
import joblib
import pandas as pd
from huggingface_hub import hf_hub_download
path = hf_hub_download("agenthinkmesh/gcc_education_ai", "model_dropout_risk.pkl")
model = joblib.load(path)
X = pd.DataFrame([{
"uni_enc": 6, # Kuwait University
"faculty": 3,
"year_of_study": 2,
"entry_gpa": 80.0,
"current_gpa": 58.0,
"is_female": 1,
"is_kuwaiti": 1,
"scholarship": 0,
"works_part_time": 0,
"commute_hours": 0.5,
"attendance_pct": 0.55,
"failed_courses": 3,
"family_pressure": 3,
"financial_stress": 4,
"social_integration": 2,
"mental_health_score": 45.0,
"language_barrier": 0,
"wrong_major": 1,
"advisor_meetings": 1,
"oil_job_pull": 0,
}])[list(model.feature_names_in_)]
print(model.predict(X)) # 1 = predicted dropout
print(model.predict_proba(X)) # [P(stay), P(dropout)]
automl_results.json lists the expected columns for each model under <model_key>.features. That list does not apply to model_arabic_proficiency.pkl or model_special_needs.pkl, as noted above.
Rerunning the training pipeline (this also regenerates consistent versions of the two mismatched pickles):
pip install numpy pandas matplotlib scikit-learn==1.8.0 joblib
# The script writes to OUT = "/home/claude/gcc_education_ai" (line 39); edit that path first.
python gcc_education_ai.py
The script writes all ten .pkl files, automl_results.json, the dashboard PNG and a short README.md to OUT. If you point it at a clone of this repo, that README.md replaces this card. Exact numbers can vary slightly across NumPy and scikit-learn versions.
Intended use
- Demonstrating and teaching a tabular AutoML workflow on GCC education themes.
- Prototyping dashboards and what-if tools before real, governed data is available.
- A template for retraining on real institutional data, with proper ethics review.
Limitations and risks
Synthetic data only. The relationships are hand-written by the script author. They have not been validated against any real student, teacher, school or ministry data.
Do not use for decisions about individuals. Admissions, scholarships, dropout interventions, staffing and special-needs provision are all out of scope.
Sensitive attributes. Several inputs are protected or sensitive:
- gender (
is_female) - nationality (
is_kuwaiti) - disability or orphan status
consanguinity_rate- mental-health scores
Some synthetic labels depend on these directly. The scholarship label, for example, adds weight for
is_kuwaiti. Any real-world version needs fairness and legal review.- gender (
Low-performing model. Special Needs Support Demand reaches only R² 0.257 on the test set, according to the JSON. Its uploaded pickle is also not the model those metrics describe.
Mismatched artifacts.
model_arabic_proficiency.pklandmodel_special_needs.pkldo not match the code or metadata. Their inputs are undocumented.Accuracy compared with baselines. Positive rates are low for several tasks, so compare accuracy with the majority-class baseline:
- Dropout Risk: about 0.82
- Teacher Retention: about 0.78
- STEM Readiness and Scholarship: about 0.65
- EdTech Adoption: about 0.55
Only accuracy is reported. There is no precision, recall or calibration.
Pickle security.
.pklfiles can execute code when loaded. Only load files you trust.
License
MIT
Citation
@misc{agenthink2026gcceducation,
author = {AgenThink},
title = {GCC Education AI: 10-Model Suite},
year = {2026},
publisher = {Hugging Face},
howpublished = {\url{https://huggingface.co/agenthinkmesh/gcc_education_ai}}
}
- Downloads last month
- -