Distill training pairs for SHAP narrator

[William-Hill]

Summary

Generate training pairs for the new SHAP narrator task. Each pair maps a student's SHAP values + profile to an advisor-facing narrative with grounded explanations and interventions.

Depends On

• Add SHAP narrator task type to training pipeline #97 (SHAP narrator task type must exist in pipeline)

Prerequisites

• SHAP data must be populated in student_level_with_predictions.shap_explanations (run ML pipeline with SHAP step)
• Readiness scores must exist in llm_recommendations (run readiness score generator)

Tasks

• Ensure SHAP data is populated in DB (run python ai_model/complete_ml_pipeline.py if needed)
• Run python -m training.distill --school bishop-state --task narrator (~1,500 pairs)
• Run python -m training.prepare --school bishop-state --task narrator
• Verify SHAP grounding: spot-check 10 pairs — do narratives reference actual SHAP features?
• Commit training data to training_data/bishop-state/
• Track distillation cost (expected: $2-4)

Acceptance Criteria

• = 1,200 validated narrator pairs after dedup

• Train/val/test splits at training_data/bishop-state/final/narrator/{split}.jsonl
• Spot-check confirms narratives cite specific SHAP features by name and magnitude