Teaching a Small LLM Scholastic Voice

Fine-Tuning Qwen 2.5 on the Catechism, Summa, and Augustine via Local MLX

Pablo Leyva · R1 New Jersey Institute of Technology

Read the paper View the poster Try the model

This page hosts the latest rendered artifacts for the scholastic-llm project. Both PDFs are regenerated by GitHub Actions (paper, poster) on every push that touches their sources.

The project is complete through Phase 2. Four MLX LoRA adapter variants are published on Hugging Face, with sft-v2-iter400 as the best checkpoint (68/120 strict, 68/90 balanced). The DPO chain is documented as a negative result.

Try the model

Live interactive demo on Hugging Face Spaces (free CPU tier — 2 vCPU, 16 GB RAM). The 7B model is large for this hardware: each response takes ~5 minutes (responses are capped at 220 tokens; full Summa-form answers are in the paper). For real-time inference, run the MLX adapter locally on Apple Silicon (a single response in ~1-3 seconds).

⚠ Research experiment, not theological authority. The model described in this work is a stylistic and pedagogical experiment. Its outputs are not authoritative on Catholic doctrine, biblical interpretation, or theology. They must not be cited as catechetical instruction. For doctrinal questions, consult the actual Catechism of the Catholic Church, a qualified priest, or a trained theologian.

Repository: github.com/pleyva2004/scholastic-llm