amanrr48/prajna-mistake-predictor

PRAJNA Mistake Predictor

Logistic regression model that predicts P(mistake) per student per micro-topic for NEET/JEE exams.

Architecture

• Model: Scikit-learn LogisticRegression with StandardScaler
• 7 features: rolling_accuracy, ability_score, topic_difficulty, exam_importance, avg_time_spent, streak, exam_number
• Label: binary (1 = student accuracy < 50% on topic in exam, 0 = OK)
• Training: Exams 1-8, validation on 9, test on 10

Features

Feature	Description	Range
rolling_accuracy	Avg accuracy on this topic across prior exams	0-1
ability_score	Student's subject-level baseline ability	0-1
topic_difficulty	Avg difficulty from exam DB, normalized	0-1
exam_importance	PRAJNA appearance probability for this topic	0-1
avg_time_spent	Normalized avg time on this topic	0-1
streak	Consecutive correct/wrong streak (clipped)	-1 to 1
exam_number	Temporal position normalized	0-1

Outputs

• p_mistake: probability the student will make a mistake on this topic (0-1)
• feature_importances: which factors most predict mistakes

Usage

from mistake_predictor import MistakePredictor

mp = MistakePredictor()
X_train, y_train = mp.build_features(results_df, abilities_df, topic_difficulty, prajna_importance)
mp.train(X_train, y_train)

# Per-student predictions
predictions = mp.predict_for_student(results_df, abilities_df, topic_difficulty, prajna_importance, student_id="S1")

Center-View Analysis

The companion MistakeAnalyzer provides aggregate insights:

• Danger Zones: topics where error_rate > threshold AND PRAJNA appearance_probability is high
• Co-failure Patterns: P(fail B | fail A) -- correlated topic failures
• Time vs Accuracy: scatter analysis revealing conceptual gaps

License

Apache 2.0