asandeistefan/romanian-baccalaureate-mathematics

Romanian Baccalaureate in Mathematics

A curated collection of Romanian Baccalaureate (BAC) mathematics examination papers and answer keys, transcribed from PDF to structured Markdown using Vision-Language Model OCR. Currently the years 2019 - 2025 were added, more will be processed soon.

Directory Structure

romanian-baccalaureate-mathematics/
├── metadata.csv          # Index of all exam papers
├── pdfs/                 # Original PDF files
│   ├── 2023_matematica_m1_subiect.pdf
│   └── ...
└── markdown/             # OCR-processed content
    ├── 2023_matematica_m1_subiect.md
    └── ...

Metadata CSV Schema

metadata.csv contains one row per exam paper (questions only). Answer sheets are linked via reference rather than duplicate entries.

Column	Type	Description
filename	string	PDF filename (e.g., 2023_Matematică_M_științe_subiect.pdf)
year	integer	Academic year of the exam
subject	string	Subject code (e.g., Matematică)
profile	string	Educational track (e.g., Real, Uman, Pedagogic) or empty if not specified
answer_sheet	string	Filename of corresponding answer key/barem, or empty string if unavailable

Note: Rows where answer_sheet is empty indicate that no official answer key was published or available in the source API for that particular exam variant.

Markdown Format

Each .md file follows a normalized structure:

• Headers: Section headers marked with ## SUBIECTUL I, ## SUBIECTUL al II-lea, etc.
• Math: All mathematical notation in LaTeX ($...$ for inline, $$...$$ for display)
• Formatting:
  • Exercise labels in bold (e.g., **1.**, **a)**)
  • Point values preserved (e.g., 5p, 3p)
  • Tables converted to plain text with pipe characters removed
  • Boilerplate headers/footers stripped (Ministry logos, page numbers, etc.)

Data Processing Pipeline

1. Ingestion: Documents fetched via API from official Romanian Ministry of Education repositories
2. OCR: PDF pages converted to images (300 DPI) and processed through Vision-Language Model (Gemini/Hunyuan) with prompts optimized for Romanian mathematical notation
3. Cleaning:
   • Removal of running headers/footers
   • Deduplication of headers across multi-page documents
   • Table flattening and normalization
   • Consistent double-spacing between content blocks

Usage

Loading with Hugging Face Datasets

from datasets import load_dataset
import pandas as pd

ds = load_dataset("asandeistefan/romanian-baccalaureate-mathematics", split="train")

df = pd.read_csv("metadata.csv")
with open(f"markdown/{df.iloc[0]['filename'].replace('.pdf', '.md')}") as f:
    content = f.read()

Loading with Pandas

import pandas as pd

df = pd.read_csv("metadata.csv")
print(f"Total exams: {len(df)}")
print(f"Years covered: {df['year'].min()} - {df['year'].max()}")

Limitations

• OCR Artifacts: Despite cleaning, occasional transcription errors in complex geometric diagrams or handwritten annotations may persist
• Missing Answer Sheets: Some exam variants lack official answer keys in the source database; these entries have empty answer_sheet values

Citation

@dataset{romanian_bac_math_2026,
  title = {Romanian Baccalaureate in Mathematics Dataset},
  author = {Asandei Stefan-Alexandru},
  year = {20246,
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/username/romanian-baccalaureate-mathematics}
}