Building Evals for a CPT-Tuned Local LLM: Exact, Partial and Semantic Matching

funJS · reddit · 2026-09-24

A developer shares how to build evals for a small model (Qwen 4B) after continued pretraining (CPT) on a new domain — the model outputs n-transfer subway journey legs. He constrains output with a strict JSON schema, then fine-tunes post-CPT via Unsloth SFT (alpaca format) to improve schema reliability.

Beyond exact matches, evals also measure partial matches (e.g., predicting part of a route) and semantic matches, together assessing whether knowledge was truly internalized. The author reports the method works well; details on his blog.

Related event: Practical Guide: Designing Evals for Continually Pretrained Small Models(4 posts)→

Original post →

More from coding & agent

coding & agent channel →