feat: config-driven distillation pipeline for per-school fine-tuned models#93

Open

William-Hill wants to merge 18 commits intomainfrom

feature/distillation-pipeline

Commits on Feb 24, 2026

docs: design doc for self-service data upload (issue #86)
William-Hill
committed
docs: implementation plan for self-service data upload (issue #86)
William-Hill
committed

Commits on Mar 28, 2026

docs: design spec for config-driven distillation pipeline
William-Hill
committed
docs: implementation plan for config-driven distillation pipeline
William-Hill
committed
chore: scaffold training pipeline package and test infrastructure
William-Hill
committed
feat(training): config loader with YAML validation and JSONL writer
William-Hill
committed
feat(training): add Bishop State school config and seed queries
William-Hill
committed
feat(training): teacher prompt templates for explainer and summarizer
William-Hill
committed
feat(training): seed data generation for explainer and summarizer
William-Hill
committed
feat(training): distillation pipeline with dual teacher backend support
William-Hill
committed
feat(training): dataset preparation — filter, dedup, and split
William-Hill
committed
feat(training): eval harness with ship criteria for model quality gates
William-Hill
committed
feat(training): MLX QLoRA fine-tuning wrapper
William-Hill
committed
feat(training): Ollama model export and registration
William-Hill
committed
feat(dashboard): model client adapter for Ollama/OpenAI routing
William-Hill
committed
feat(dashboard): route explain-pairing and query-summary through model client
William-Hill
committed
refactor: simplify training pipeline — deduplicate code, extract shared utilities
William-Hill
committed
fix: address CodeRabbit review findings
William-Hill
committed