FiveC/vinid-results

ViNID results

Data and system outputs of ViNID-1000, a contrastive benchmark for Vietnamese name-spelling fidelity in machine translation: 1,000 pairs (2,000 attested spellings) that share one ASCII form and differ only in diacritics, evaluated on ten systems. Code and manuscript are distributed separately; this repository holds only data and the numbers they produce.

data/                     pairs.jsonl, templates.json, case-normalization.json
outputs/inputs/           E1, E2, E3 and masked inputs built from data/
outputs/frequency/        UIT-ViNames and Wikidata count evidence per pair
outputs/responses/<system>/<E>/<condition>/   raw system outputs (index.json lists all cells)
outputs/scores/<system>/<E>/<condition>/      scored records and metrics
outputs/analysis/<study>/ numeric source of every reported table and number
data-card.md              composition, evidence, rights and intended use
reproduction.md           which tool produced which analysis

Download with a fixed revision. Earlier layouts (200-pair studies, previous benchmark versions, working files) were removed from the current tree on 2026-11-08 and remain in the commit history, for example at revision bc91bc91312470ae693df941afe04f144c81aac2. Model weights, tokenizers and credentials are not stored here; source-audit records that map pairs to roster evidence are kept offline.