Datasets

Open Dataset Library

We release all of the raw data used to measure and validate the AIO Framework standard. Model responses, prompt sets, the vocabulary, and analysis scripts — anyone can reproduce and extend them.

Principles

Open, Reproducible, Standard Format

AIO's measurement data is released not just as result charts but down to the level of individual raw responses. It is designed so that anyone can reproduce the analysis with the same data, add a new model, or evaluate a new domain.

Currently Available

Open Datasets

AIO Framework-Bench-v1·366,120 responses·8 models

AIO 20003 v1.0 raw data

7 domains (MED / BIZ / TECH / EDU / LAW / DEF / CARE), 8 frontier models, 366,120 responses in total. Each response includes a AIO Framework log, vocabulary-matching results, and domain metadata.

Format: JSONL · License: CC BY 4.0 · Size ≈ TBD

Download URL TBD · 공식 호스팅 확정 후 게시

Vocabulary-v1·39 codes

AIO 00011 vocabulary v2.0

A machine-readable vocabulary of 39 codes: Schwartz 19 + Walton 10 + Hovland-Kelley 10. It includes the codes, their Korean and English names, definitions, and referenced academic sources.

Format: JSON / CSV · License: CC BY 4.0 · GitHub repository coming soon

Prompts-v1·14,175 / layer

Value-Conflict Prompt Set

A collection of prompts that qualitatively design value-conflict scenarios by domain. Each prompt is tagged with its intended value conflict (e.g., autonomy vs. achievement) and a domain tag.

Format: JSONL · License: CC BY 4.0