README.md

metadata

pretty_name: EduPlanBench Prepared Data
language:
  - en
task_categories:
  - question-answering
  - text-generation
  - reinforcement-learning
tags:
  - education
  - benchmark
  - planning
  - knowledge-tracing
  - tutoring
license: other

EduPlanBench Prepared Data

This artifact contains the prepared EduPlanBench data used by the benchmark code. It is intended to be downloaded into the repository's data/ directory so experiments can run without rebuilding from raw datasets.

Prepared data is available in two access patterns:

• Use the Hugging Face branch/revision dropdown to switch between 10k and 35k.
• Stay on main and download from versions/10k/ or versions/35k/.

Contents

processed/
  track1_text_math/
  track2_mooc_planning/
  track3_kt_simulator/
tasks/
  track1_text_math/
  track2_mooc_planning/
  track3_kt_simulator/

Prepared versions:

Version	Build command	Notes
10k	build-tasks --track all --limit 10000	Track 1/2/3 each have 10000 task instances.
35k	build-tasks --track all --limit 35000	Track 1 has 35000, Track 2 has 35000, Track 3 has 33397 XES3G5M-derived task instances.

Version counts:

Version	Track 1	Track 2	Track 3
10k	10000	10000	10000
35k	35000	35000	33397

Use In EduPlanBench

From the EduPlanBench repository root:

python3 scripts/download_prepared_data_from_hf.py \
  --repo-id erwinmsmith/EduPlanBench-data \
  --version 10k \
  --data-dir data

Use --version 35k to download the larger prepared task bank.

To download from a Hugging Face branch/revision root instead:

python3 scripts/download_prepared_data_from_hf.py \
  --repo-id erwinmsmith/EduPlanBench-data \
  --revision 35k \
  --path-in-repo . \
  --data-dir data

Then run experiments directly:

python3 -m eduplanbench experiment \
  --tracks all \
  --agents react,one_shot,step_by_step,cot \
  --llm deepseek \
  --limit 300 \
  --sample random \
  --sample-seed 42

Source Data

The prepared files are derived from Eedi public data, MathDial, MisstepMath, MOOCCubeX, XES3G5M, and EdNet/KT1-style data. Raw datasets are not included here. Follow the upstream licenses and access rules for each original dataset. The repository file rawdataset/README.md documents the expected raw layout and source download locations for rebuilding from scratch.

Rebuild From Raw Data

If all raw datasets are placed under rawdataset/:

python3 -m eduplanbench data prepare --track all
python3 -m eduplanbench build-tasks --track all --limit 10000