Constructing a Supervised Clinical Terminology Corpus from SNOMED CT RF2 Files: Five-Version Iterative Design, BFS Bottom-Up Hierarchy Processing, and Empirical Validation through QLoRA Fine-Tuning

We describe the iterative construction of a 2,000,000-example supervised corpus for clinical terminology alignment, derived exclusively from SNOMED CT Release Format 2 (RF2) files. The corpus development proceeded through five successive versions, each motivated by specific failures identified in downstream QLoRA fine-tuning experiments on Mistral-7B-Instruct-v0.2. Version 1 formulated the task as code prediction and was abandoned after systematic gradient norm spikes at epoch position 0.154–0.160 across all training sessions, independent of learning rate. Versions 2 through 5 progressively introduced five semantic example types: canonical FSN definitions with synonym enumeration (Type 1), synonym-to-FSN equivalence mappings (Type 2), clinical text definitions with adaptive length filtering (Type 3), hierarchical is-a chain instances via bottom-up iterative BFS (Type 4), and SNOMED model attribute-value pairs (Type 5). A key technical contribution is the replacement of recursive hierarchy traversal—which caused StackOverflow on the 350,000+ node SNOMED CT graph—with a completely iterative BFS algorithm that processes each concept only after all its ancestors have been resolved. Statistical analysis of the final corpus (v5_final) confirms low length variance: mean 67.8 tokens, 95th percentile 121 tokens, P95/mean ratio 1.78—comparable to standard NLP benchmarks. Empirical validation through six QLoRA training runs produced minimum cross-entropy loss of 0.370 and maximum token accuracy of 91.5% on the final corpus design, versus loss 0.377 and accuracy 91.3% on corpus v2 (44,000 examples), demonstrating measurable improvement attributable to corpus design decisions.