pisa_multimodal_dataset_card.md

# Dataset Card: PISA Multimodal (Parallel & Not-Parallel)

## Summary
This dataset contains **48 parallel multimodal samples** (paired TXT↔PDF) derived from **PISA studies up to 2012**, plus **47 non-parallel samples** (TXT-only or PDF-only). Each sample may include multiple questions. Content is available in **German** and **English**.

**Source & usage:** Materials are published by the **OECD** and are provided here **for non-commercial use** only. Please verify that your usage complies with OECD terms.

**Quality note:** Some of the **parallel** samples are **noisy** (e.g., OCR artifacts or imperfect alignments). These are **explicitly labeled in the filename** so they can be filtered out during experiments.

## Organization & Splits
- **`train`** — *parallel* split: 48 samples (TXT and matching PDF).
- **`not_parallel`** — 47 samples where only TXT **or** only PDF is present.  
  *Note:* In this split, one of `Content` or `PDF` can be `null`.

*(Counts reflect the provided corpus; they may change if you add/remove files.)*

## Languages
- **German:** `deu_Latn`  
- **English:** `eng_Latn`

## Tasks
- **Reading** comprehension  
- **Math** problem solving

## Data Fields (Columns)
- **`ID`** *(int32)* — Running index of the sample (per build script).
- **`Task Name`** *(string)* — Base name of the material (filename stem before the last underscore), e.g., `Bookshelves`.
- **`PDF`** *(Pdf feature)* — Full PDF object (stored via `datasets.Pdf`). May be `null` in `not_parallel` if missing.
- **`Language`** *(string)* — Language code of the folder, e.g., `deu_Latn` or `eng_Latn`.
- **`Task`** *(string)* — Task category from the folder name: `reading` or `math`.
- **`Content`** *(string)* — Full text from the paired `.txt`. May be `null` in `not_parallel` if missing.
- **`TXT Filename`** *(string)* — Basename of the associated `.txt` file (if present).
- **`PDF Filename`** *(string)* — Basename of the associated `.pdf` file (if present).

## Source Data & Provenance
- **Origin:** Items compiled from **OECD PISA** materials (≤2012).  
- **Structure:** Files are organized by `<language>_<task>` folders, e.g., `german_reading`, `english_math`.  
- **Derivation:** Parallel pairs were formed by matching TXT and PDF stems; the `not_parallel` split contains items missing one counterpart.

## Licensing & Usage
- **License:** *Other / Non-commercial.* The content is published by the **OECD**. Redistribution and use should comply with **OECD non-commercial terms**. This card is not legal advice—please review the official terms for your use case.

## How to Load
```python
from datasets import load_dataset

# From the Hub (replace with your repo id)
ds = load_dataset("your-username/your-dataset")

# Parallel split
train = ds["train"]

# Not-parallel split
not_parallel = ds["not_parallel"]

# Example access
row = train[0]
print(row["Task Name"], row["Language"], row["Task"])
# PDF is a Pdf feature record with a path/bytes:
print(row["PDF"])
```

## Intended Uses & Limitations
- **Intended:** research on multimodal QA, OCR+NLP, cross-lingual evaluation, dataset tooling.  
- **Not intended:** commercial exploitation of OECD content; deployment without verifying rights.  
- **Limitations:** Some samples include multiple questions per PDF/TXT; `not_parallel` items are incomplete by design; some parallel items are flagged as **noisy** in filenames.

## Citation
If you use this dataset, please cite the OECD PISA materials appropriately and reference this dataset card.