README.md

CodiEsp Diagnosis Grounding Data

This repository contains the prepared data files needed to reproduce the CodiEsp-D diagnosis grounding experiments. It is a data artifact only: it does not include experiment code, Slurm scripts, model outputs, logs, or raw source archives.

Directory Layout

data/codiesp/
  facts_test_full_exact.jsonl
  evidence_relocations_full_exact.jsonl
  stats_full_exact.json
  test_docs_full_exact.txt
  spotcheck_50_full_exact.tsv

index/icd10cm_fy2018/
  icd10cm_fy2018_retrieval.jsonl
  code_metadata.jsonl
  inventory_manifest.json
  self_retrieval_probe.json

schema/icd10cm/
  normalization_map.json
  vocab_family.json
  vocab_role.json
  vocab_qualifier.json
  vocab_scope.json
  vocab_temporal.json
  vocab_temporal_raw.json

Main Evaluation File

data/codiesp/facts_test_full_exact.jsonl

This is the prepared CodiEsp-D test set used by the second-domain experiments. It contains the exact-English-relocation slice of the official CodiEsp test split.

Summary from data/codiesp/stats_full_exact.json:

• Source documents: 250
• Target facts: 3,144
• Unique gold diagnosis codes: 958
• Exact relocation parse rate: 1.0
• Exact substring rate: 1.0

Each JSONL row contains the clinical mention, English context fields, document identifier, one gold diagnosis code, and relocation metadata.

Relocation Files

data/codiesp/evidence_relocations_full_exact.jsonl
data/codiesp/spotcheck_50_full_exact.tsv

These files document how Spanish CodiEsp diagnosis spans were relocated into the English context used for retrieval. The evaluation JSONL keeps only facts with exact English substring relocation.

Retrieval Inventory

index/icd10cm_fy2018/icd10cm_fy2018_retrieval.jsonl

This is the candidate label inventory used by retrieval. It contains 71,344 candidate diagnosis-code entries. Each entry contains a code identifier, canonical English label, retrieval/documentation text, and structural metadata.

Inventory construction metadata is in:

index/icd10cm_fy2018/inventory_manifest.json

Schema Files

schema/icd10cm/

This directory contains the controlled vocabularies and normalization map used by the CodiEsp structured grounding schema.

Not Included

The following are intentionally excluded:

• Experiment code and Slurm scripts.
• Model outputs and intermediate run files.
• Slurm logs.
• Raw downloaded CodiEsp or diagnosis-code archives.
• Python bytecode caches.

MANIFEST.tsv lists the files in this data artifact with byte sizes.