Current outcome: On 12 August 2026, the provider told me that it had replaced the file. Korea's Public Data Portal now hosts the 601-row
한국한의약진흥원_한약재 정보_20260811dataset. Compared with the old version, 333 cells across 270 records have changed. The revision was extensive, but some problems remain.
I noticed an odd spelling in the row for Terminalia fruit, identified as KMH2. The dataset description said Terminalia Fruit, but both the preview and the downloadable CSV said Termilia Fruit. The missing characters were na.
The same shape appeared in other rows. Gentiana was written as Gentia, Cinnamomum as Cinmomum, and officinale as officile. The repeated loss of na across unrelated names and fields suggested that these might not be isolated typos.
The values were wrong in the CSV, not just on screen
I first preserved the public CSV as it was. Excluding the header, it contained 601 rows and seven columns, and none of the 601 herbal-medicine IDs was duplicated. I then compared every field in the first 50 rows with the portal's official preview. They all matched, so the spellings were in the distributed file rather than being a display glitch.
In the first sample, I found the case-insensitive loss of na in 32 cells across 16 rows. I checked part of that sample against Korean pharmacopoeial material and Kew plant-name resources. Values that I could not settle remained provisional. Running the same rule across the full file narrowed the manual review queue to 88 rows.
Those 88 rows were candidates, not 88 confirmed errors, and the number cannot be used as an error rate. I did not insert na automatically and treat the result as corrected data. I wrote to the provider instead, asking it to check the source records and review all 601 rows.
The hypothesis and evidence I sent
My request proposed one pipeline hypothesis: a setting meant to transform the missing-value token NA might also have acted on lowercase na inside ordinary words. I included enough information to identify the original file and supplied examples by row. I also described the effects on exact-name searches and text joins and attached the script used to reproduce the screening.
The provider replied the next day that it had found an error in the missing-value conversion setting. It said it had reviewed the English-name, Latin-name, and scientific-name fields in all 601 rows, corrected the affected values, and replaced the CSV on the Public Data Portal. The explanation about the setting comes from the provider's reply; the two public files cannot independently reveal its internal pipeline. I used the private reply only to attribute that explanation and have omitted the staff member's name and the correspondence itself.
What actually changed in the replacement
The old and current files both have 601 rows, seven columns, and the same set of IDs. A cell-by-cell comparison found 333 changed values in 270 records. Of those changes, 18 were in the English-name field, 56 in the Latin-name field, and 259 in the scientific-name list. Each of the 88 candidate records changed in at least one field. Among the visible repairs were Termilia Fruit to Terminalia Fruit and Zingiber officile to Zingiber officinale.
The provider repaired 30 of the 32 original suspect cells. Two Gentia lutea entries remain, and the replacement introduced 163 literal question marks across 150 records. I cannot tell from the public files how those characters appeared.
Verified result and remaining limits
I preserved the original CSV, traced the repeated loss of the same string, and submitted a correction request with evidence by row. I based the request on the match between the CSV and official preview. The recurrence of the missing na across taxa and the checks against naming sources provided further support.
The provider's reply and the replacement file are the external record of the correction. The official portal hosts the 11 August 2026 revision, and a direct comparison with the old version shows changes in 270 records. The final state is a replaced file with major repairs, not a fully resolved dataset. Some spellings remain suspect, and the question marks are new. The figure 601 is the total number of records. It is neither an error count nor the number of changed rows.
The damaged strings could cause failures in exact-name searches, links to external taxonomy identifiers, and straightforward text joins. Those are the interoperability effects I could test directly. I found no basis for a claim of clinical harm or for describing the correction as proof of improved patient safety.
The turning point was checking whether several odd spellings followed one rule. A script reduced the number of rows that needed manual attention, and the provider could then compare those candidates with its source data. Rechecking the replacement kept confirmed repairs separate from the problems that remain.
Public evidence and reproduction method
- Official dataset on Korea’s Public Data Portal
- Dataset ID:
15109111 - Current public version:
한국한의약진흥원_한약재 정보_20260811 - Current public size: 601 rows and seven fields
- Provider: National Institute for Korean Medicine Development
I checked the CSV structure before running the string-pattern script. Rows that needed manual review went into a separate evidence table. After the replacement, I compared the two files cell by cell. I used NCBI Taxonomy and GBIF to find candidates and test identifier links. I did not use them as the final authorities for the names.