0.2.0

AIED-Unplugged Dataset SDK

Python SDK for the AIED Preview Competition: the official graders, dataset loaders, exploration helpers, and scaffolding to run a Hugging Face model on a track.

pip install aied-unplugged            # graders, loaders, submission tooling
pip install 'aied-unplugged[all]'     # + kagglehub, datasets, matplotlib, transformers

The SDK is optional. It reads the published dataset and writes the published submission format.

---

The tracks

---

1. Grading

from aied_unplugged import evaluate

result = evaluate("math", "my_predictions.csv", "validation_reference.csv")
result.primary  # 0.41…  the ranking metric
result.metrics  # every metric the track reports
result.per_example  # one row per item: true, pred, correct

References can be a CSV or a validation split loaded from the dataset, with the same target columns either way.

from aied_unplugged import load_track, evaluate

validation = load_track("math", "validation")
evaluate("math", predictions, validation)

Metrics

A constant prediction zeroes the QWK expected-agreement denominator. That case scores 1.0 when the prediction is right everywhere and 0.0 otherwise, instead of NaN. Macro F1 skips classes absent from the reference: averaging over the full taxonomy hands you a guaranteed zero for classes the split never uses, and the preview sample omits two of the thirteen.

Validation checks

A grader rejects the whole file instead of scoring the rows it can read. It raises SubmissionError naming the ids and columns at fault, without quoting a target value. Rejected files include:

• an empty file, a missing column, or an extra column
• a blank or duplicated identifier
• an empty prediction
• an ungraded row, or a missing row
• a competence score off the 0/40/80/120/160/200 grid
• a diagnostic outside the thirteen labels
• an answer-sheet value outside the seven
• a question_number that is not an integer (1.0 and "1" both fail)
• a question repeated, missing, or not on the sheet

---

2. Loading

Getting the dataset

download() pulls the dataset from Kaggle. Needs aied-unplugged[kaggle].

from aied_unplugged import download, load_track

download()  # into the kagglehub cache, returns the path
load_track("math")  # the loaders below now find it
load_track("math", source="kaggle")  # same, in one call
load_track("math", source="hf")  # in-memory DatasetDict from the HF Hub

Later calls reuse the cache; pass force=True to re-fetch. download() also sets the default root for the process, so explore and verify run without root=.

Both mirrors carry the same release:

• Kaggle: https://www.kaggle.com/datasets/aibox-lab/aied-unplugged-preview
• Hugging Face: https://huggingface.co/datasets/aiboxlab/aied-unplugged-preview

To unpack one by hand, put metadata/ and schema/ under a single directory, then name that directory with root=, with AIED_UNPLUGGED_DATA, or as competition-dataset/ in your working directory. root= wins over AIED_UNPLUGGED_DATA, which wins over download(), which wins over competition-dataset/.

Loading a track

from aied_unplugged import load_track, load_metadata, open_image

splits = load_track("essays")  # local release tree if present, else the Hub
splits.train, splits.validation, splits.test

train = load_track("math", "train")  # one split
image = open_image(train.iloc[0], "equation_image")

The local loaders return pandas DataFrames with the Parquet metadata and absolute image paths in <column>_path.

load_schema("math")  # the published taxonomy, fields and metrics
graded_ids("aes")  # the ids a submission must carry
question_counts()  # questions per graded answer sheet
sample_submission("aes")  # the published placeholder file
verify()  # re-check every SHA-256 in the release

---

3. Exploring

Needs aied-unplugged[explore] and a local copy of the dataset, from download() or from a mirror you unpacked yourself.

from aied_unplugged import explore

explore.summary()  # rows per track and split
explore.label_distribution("math")  # counts and shares, taxonomy order
explore.plot_label_distribution("math")
explore.describe_taxonomy("math")  # labels with their Portuguese source terms

explore.show_essay("essay-1-3")  # the page, with its five competence scores
explore.show_math("math-604-01")  # question above, student working below
explore.show_sheet(0)  # a sheet with what was marked
explore.render_math("math-604-01")  # the same pair as one PIL image

Name an item by id, by position, or by passing its metadata row.

---

4. Running a model

Needs aied-unplugged[models]. Two strategies, either one enough for a valid submission and a baseline score.

from aied_unplugged import models, submission

# Predict the training majority for everything.
predictions = models.majority_baseline("math")

# Zero-shot with a local vision-language model.
predictions = models.predict("math", "Qwen/Qwen2.5-VL-3B-Instruct", limit=20)

submission.write("math", predictions, "submission.csv")

predict prompts the model once per item with the track's images and parses the reply into the submission schema. An unparseable reply falls back to a safe default, so a run always ends with a gradeable file.

Fine-tuning covers the math track only, as image classification over the student's working:

trainer = models.finetune(model="google/vit-base-patch16-224-in21k", epochs=3)
predictions = models.predict_with_classifier(trainer, split="test")

Not included: essay or answer-sheet fine-tuning, multi-GPU, LoRA, hyperparameter search.

---

5. Submissions

from aied_unplugged import build, graded_ids, question_counts, validate, write

frame = build("answer-sheet", [{"sheet_id": "s1", "answers": {1: "A", 2: "Blank"}}])
report = validate(
    frame,
    "answer-sheet",
    expected_ids=graded_ids("answer-sheet"),
    expected_questions=question_counts(),
)
if not report.valid:
    print(report)
write("answer-sheet", frame, "submission.csv")

validate checks the exact column set, the identifiers, the competence scale, the diagnostic labels, and the JSON encoding of the answer-sheet column, including integer question numbers. With expected_questions, it also checks the question count on each sheet. build takes answers as a list of records or as a {question_number: label} mapping and encodes both to the same JSON.

---

Development

uv run --with-editable . --with pytest pytest -q
uv run ruff check .

Tests needing the release tree skip when competition-dataset/ is absent.

MIT licensed. The dataset itself is CC BY 4.0 and is published separately.