Generate a 500-concept SNOMED-based dataset with humanized (LM-adapted) terms

[filiperochalopes]

User story

As a researcher, I want a dataset of 500 randomized active SNOMED CT concepts with a more natural, human-friendly phrasing so we can test mapping and matching workflows using less technical input terms.

Use case

Sample 500 SNOMED concepts (proportionally by semantic tag), then use an LM to rewrite the terms into a more natural phrasing while preserving meaning.

Requirements

• Generate a reproducible random sample of 500 active SNOMED CT concepts from the RF2 release ZIP:
  • Use fixed seed (42)
  • Proportional sampling by semantic tag (default: disorder + finding)
  • Output CSV containing concept_id, fsn, preferred_term, semantic_tag
• Produce a second dataset version where each term is adapted by an LM to be more humanized/natural and less technical.
• Keep both versions available for comparison (original vs humanized) using the same 500 concept_ids.

Acceptance criteria

• A CSV exists with 500 sampled active SNOMED CT concepts (concept_id, fsn, preferred_term, semantic_tag) generated reproducibly from the RF2 ZIP using the provided sampling logic.
• A second CSV exists for the same 500 concept_ids containing LM-humanized terms aligned to the original entries.
• The humanized output is less technical while preserving the original clinical meaning.

[filiperochalopes]

@paynejd dataset is ready

snomed_disorder_finding_sample_20251001_500.csv

[ocl-automation]

Automated Triage — 2026-04-14

Dimension	Score
Complexity	3
Scope	2
Clarity	4
Dependency	2
Risk	2

Decision: decision/ai-ready
Type: type/feature
Signals: signal/low-risk, signal/small-scope, signal/well-specified

Moderate complexity involving SNOMED data processing, sampling algorithms, and LM integration across a few files. Well-defined requirements with clear acceptance criteria. Low-moderate risk since incorrect sampling or term generation affects research quality but doesn't break production systems.

---

Generated by ticket-triage pipeline v3. Model: anthropic/claude-sonnet-4 (openrouter).

[filiperochalopes]

/uat @paynejd