Modeling Student Learning with 3.8 Million Program Traces
Paper • 2510.05056 • Published
You need to agree to share your contact information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
Access to the PencilCode dataset requires manual approval from the Pencil Code team. This dataset contains anonymized interaction traces from students (including K-12 learners). By requesting access, you agree to use this data solely for non-commercial research purposes, to protect the privacy of the students whose interactions are represented, and not to attempt to re-identify any individuals. Please describe your intended research use in the field below.
Log in or Sign Up to review the conditions and access this dataset content.
PencilCode Program Traces
Paper: Modeling Student Learning with 3.8 Million Program Traces
Code: meghabyte/pencilcode-public
Contact: megha@cs.stanford.edu, alexisro@mit.edu, jjb@eng.ufl.edu, jda@mit.edu
Gated Dataset — Manual Approval Required.
Access is granted pending review by the Pencil Code team. Please submit your request above with a brief description of your intended research use. Any requests will be reviewed by Jeremiah Blanchard at jjb@eng.ufl.edu
Dataset Summary
This dataset contains 3.8 million programming reasoning traces collected from users of Pencil Code, a free open-source educational platform where students learn programming through visual block-based and text-based coding (CoffeeScript, JavaScript, HTML/CSS).
Each trace records a student's full interaction history with a single assignment — the ordered sequence of program states they wrote, from first attempt to final submission — along with associated timestamps. The dataset spans 9 years (2015–2024) and covers over 1.3 million unique anonymized users.
Unlike datasets that capture only final program submissions, PencilCode traces reveal the iterative reasoning process: exploratory behavior, debugging strategies, goal backtracking, and stylistic personalization. This makes the dataset uniquely suited for studying how people learn to code, not just what they produce.
Dataset Details
| Property | Value |
|---|---|
| Total traces | ~3.8 million |
| Unique anonymized users | ~1.3 million |
| Avg. traces per user | 2.86 |
| Total size | ~248 GB |
| Date range | 2015–2024 |
| Programming languages | CoffeeScript, JavaScript, HTML/CSS |
Data Fields
Each record contains:
username— Hashed/anonymized student identifiertitle— Assignment name (e.g.,snowman,lighthouse,confetti)programs— Temporally ordered list of program states written by the studenttimestamps— Execution timestamps corresponding to each program state
Data Splits
The dataset is organized into four evaluation splits mirroring those used in the paper:
| Split | Description |
|---|---|
seen_student_seen_title |
In-distribution: both student and title seen during training |
unseen_student_seen_title |
~259K traces; held-out students on known assignments |
seen_student_unseen_title |
~71K traces; known students on new assignments |
unseen_student_unseen_title |
~8.8K traces; fully out-of-distribution |
Intended Uses
This dataset is intended for non-commercial academic research, including:
- Training and evaluating language models for research on human code edit sequences
- Studying student learning behavior and programming skill development
- Building student modeling, knowledge tracing, and intelligent tutoring systems
- Investigating personalization and few-shot adaptation in educational AI
- Research on code generation, diversity, and style preservation
Out-of-Scope Uses
- Commercial applications of any kind
- Any attempt to re-identify individual students
- Deployment in production educational systems without further ethics review
Privacy
All student identifiers are cryptographically hashed. The dataset does not contain names, emails, IP addresses, or any other directly identifying information. Data access is gated and requires approval from the Pencil Code team specifically to protect the privacy of K-12 students whose interactions are represented.
Citation
If you use this dataset, please cite:
@inproceedings{ross2026pencilcode,
title = {Modeling Student Learning with 3.8 Million Program Traces},
author = {Ross, Alexis and Srivastava, Megha and Blanchard, Jeremiah and Andreas, Jacob},
booktitle = {Artificial Intelligence in Education (AIED 2026)},
address = {Seoul, South Korea},
year = {2026}
}
License & Access
This dataset is shared under a custom data use agreement with Pencil Code. Access is manually reviewed and granted for non-commercial research purposes only. Any requests will be reviewed by Jeremiah Blanchard at jjb@eng.ufl.edu.
- Downloads last month
- 11