Practical Guide: Designing Evals for Continually Pretrained Small Models

A developer shares how to design evals for a locally fine-tuned Qwen 3.5 4B after continual pretraining, using strict JSON Schema constraints and a three-level matching method to measure knowledge internalization on a multi-transfer metro itinerary task.

2026-09-24 ~ 2026-09-24 · 4 related posts

• Building Evals for a CPT-Tuned Local LLM: Exact, Partial and Semantic Matching — funJS · 2026-09-24
• Building evals for a domain-tuned 4B local LLM: JSON schema plus exact, partial, and semantic matching — funJS · 2026-09-24

2 near-duplicate retellings: funJS · funJS