README.md

---
license: cc-by-4.0
task_categories:
  - text-classification
  - sentence-similarity
  - zero-shot-classification
language:
  - en
  - ru
  - sv
size_categories:
  - 10K<n<100K
tags:
  - education
  - labor-market
  - occupational-classification
  - multilingual
  - benchmark
  - cross-national
  - temporal-drift
  - embedding-evaluation
pretty_name: "ADS: A Multilingual Educational Benchmark with Cross-Market Evaluation"
---

# ADS: A Multilingual Educational Benchmark with Cross-Market Evaluation

## Dataset Description

ADS is an evaluation benchmark spanning 40,552 records from 14 universities across 4 countries and 4 labor market classification systems in 3 languages (English, Russian, Swedish).

### Dataset Summary

The benchmark tests whether independently designed national classification systems converge in a shared embedding space and what that space reveals about the structure of education and work.

**Key findings:**
- Cross-system occupational convergence: Recall@1=0.718 without task-specific training
- Heterogeneous temporal drift: 44.2% zero drift, 37.0% single-direction, 18.8% multi-directional
- AI exposure prediction: R²=0.653 from frozen embeddings; job growth correctly unpredictable (R²=0.07)

### Languages

English, Russian, Swedish

### Data Instances

Each record contains a canonical ID, type (course/occupation/competency/mission), institution, title, text description, and metadata including country, language, and provenance information.

### Data Fields

- `canonical_id`: Unique identifier
- `type`: COURSE | OCCUPATION | COMPETENCY | MISSION
- `institution`: Source institution (e.g., MIT, MISIS, KTH)
- `title`: Record title
- `text`: Full text description
- `metadata`: JSON with institution, country, provenance

### Data Splits

Five benchmark tasks with deterministic train/dev/test splits:

| Task | Train | Dev | Test | Description |
|------|-------|-----|------|-------------|
| Cross-market occupation matching | varies | varies | varies | Match occupations across O*NET, ESCO, ProfStandart, SSYK |
| Cross-lingual alignment | varies | varies | varies | Align same occupations across EN/RU/SV |
| AI exposure prediction | varies | varies | varies | Predict automation exposure from text |
| Temporal drift prediction | varies | varies | varies | Predict occupation drift direction |
| Video availability prediction | varies | varies | varies | Predict multimedia availability |

### Source Data

- **Universities (14):** MIT OCW, UC Berkeley, Stanford, KTH, Chalmers, MISIS, HSE, MIPT, ITMO, MEPhI, BMSTU, Open Education (7 Russian universities)
- **Labor markets (4):** O*NET (US), ESCO (EU), ProfStandart (Russia), SSYK (Sweden)
- **Temporal versions:** O*NET 2022, 2023, 2024, 2025; ESCO v1.1.1, v1.2

## Responsible AI (RAI) Documentation

### Limitations and Constraints

- University course data is limited to publicly available syllabi and course descriptions
- Russian university data is primarily from engineering and technical institutions (MISIS, MEPhI, BMSTU)
- Swedish coverage limited to KTH and Chalmers (technical universities)
- Labor market classifications reflect national policy perspectives which may embed cultural biases

### Known or Suspected Biases

- STEM overrepresentation due to technical university focus
- English-language bias in embedding models (affects cross-lingual comparisons)
- O*NET reflects US-centric occupational structure

### Personal and Sensitive Information

- No personal data. All records are institutional/occupational descriptions
- No student data, no individual outcomes
- Course descriptions are publicly available institutional content

### Validated Use Cases

- Evaluating cross-lingual and cross-system embedding alignment
- Studying temporal drift in occupational classifications
- Benchmarking education-labor market alignment methods
- Testing what frozen text representations encode about labor markets

### Social Impact

- May inform educational policy and curriculum design
- Cross-national comparisons should be interpreted with awareness of cultural and economic context differences
- AI exposure predictions reflect current capabilities, not normative recommendations

### Synthetic Data

This dataset contains no synthetic data. All records are sourced from real institutional and governmental classification systems.

### Source Dataset Provenance

All source data is publicly available:
- MIT OCW: Creative Commons license
- O*NET: Public domain (US Department of Labor)
- ESCO: EU Open Data
- ProfStandart: Russian Ministry of Labor (public)
- SSYK: Statistics Sweden (public)
- Open Education platform courses: publicly available

### Collection and Preprocessing

- Course data collected via public APIs and web scraping of official university platforms
- Occupational data downloaded from official government sources
- Text standardized to UTF-8, deduplicated by content hash
- Embeddings computed using 5 frozen encoders (detailed in paper)

## Citation

```bibtex
@inproceedings{ads2026,
  title={ADS: A Multilingual Educational Benchmark with Cross-Market Evaluation},
  author={Anonymous},
  booktitle={NeurIPS 2026 Evaluations and Datasets Track},
  year={2026}
}
```