koutch/falcon_code

FalconCode

FalconCode is a large-scale dataset of student programming solutions, collected from multiple introductory programming courses.
It is designed for research on automatic programming feedback, code understanding, and educational AI.
The dataset has been curated and processed for SIGCSE 2024 and is described in detail in the FalconCode project page.

---

Dataset Summary

FalconCode contains programming assignments, student submissions, execution results, and rich metadata from three introductory programming courses.
It provides both the original problem statements (in Markdown) and cleaned/compilable student code.
This dataset enables research in:

• Automated feedback generation
• Code grading
• Program repair
• Code embeddings and representation learning
• Programming pedagogy

Two main subsets are provided:

• full – All submissions (including non-compiling code)
• submitted – Filtered set of submissions that compile, from standalone assignments only (i.e., those which can be automatically evaluated through unit tests)

---

Supported Tasks and Benchmarks

This dataset can be used for:

• Code Completion (fill-in-the-middle or prefix completion)
• Code Classification (e.g., correct vs. incorrect, concept tagging)
• Feedback Generation (free-text)
• Program Repair
• Educational Modeling (student modeling, performance prediction)

---

Dataset Structure

Data Fields

Each record contains:

Field	Description
problem_id	Unique problem identifier
course_id	Course identifier (2=train, 3=validation, 4=test)
source_code	Cleaned Python source code
original_source_code	Original unmodified source code
score	Unit test score (float, -1 if not executed)
concepts	List of programming concepts covered
prompt	Problem description in Markdown
type	Problem type (e.g., "lab", "project")
student_id (if present)	Anonymized student identifier
Additional metadata fields from course logs

---

Data Splits

Splits are defined by course:

Split	course_id	Description
train	2	Training data
validation	3	Validation data
test	4	Test data

Each split is further sharded into 10 parquet files for efficient streaming.

---

Processing

The raw CSVs were processed with the following pipeline:

1. Load and merge:

   • Problem statements (falconcode_v1_table_problems.csv)
   • Code samples (falconcode_v1_code_samples.csv)
   • Run results (falconcode_v1_table_runs.csv)

2. Prompt cleaning:

   • HTML problem statements converted to Markdown

3. Concept extraction:

   • Programming concept columns collapsed into a single concepts list

4. Filtering (for the "submitted" subset):

   • Removed non-standalone assignments (e.g., those requiring external files or which cannot be automatically evaluated with unit tests)
   • Removed project-type problems
   • Removed submissions with score = -1
   • Dropped rows with missing data

5. Code compilation check:

   • submitted subset contains only code that compiles

6. Code cleaning:

   • Normalized formatting and whitespace for consistency

---

Subsets

• full: All available submissions from all assignments (including non-compiling ones)
• submitted: Only submissions that compile, for standalone assignments

---

Usage

from datasets import load_dataset

# Load the "submitted" subset for training
ds = load_dataset("username/falconcode", "submitted")

print(ds["train"][0])

Citation

@inproceedings{10.1145/3545945.3569822,
author = {de Freitas, Adrian and Coffman, Joel and de Freitas, Michelle and Wilson, Justin and Weingart, Troy},
title = {FalconCode: A Multiyear Dataset of Python Code Samples from an Introductory Computer Science Course},
year = {2023},
isbn = {9781450394314},
publisher = {Association for Computing Machinery},
address = {New York, NY, USA},
url = {https://doi.org/10.1145/3545945.3569822},
doi = {10.1145/3545945.3569822},
abstract = {The lack of large and diverse datasets of student code samples limits some forms of computer science education research. To address this problem, we created FalconCode, a novel collection of over 1.5 million Python programs from over two thousand undergraduate students at the United States Air Force Academy. FalconCode captures over five semesters worth of code samples from our introduction to computing course, which is taken by every student regardless of their academic major. The dataset contains student code submissions for over 800 programming assignments, as well as additional metadata such as the prompt for each assignment, the testcase(s) used to evaluate student submissions, and the specific skills needed to solve each problem. In this paper, we describe the methodology used to create FalconCode and the steps taken to anonymize the data. We then describe FalconCode's data schema, and show how it can support a wide range of research---including those utilizing machine learning (ML) and artificial intelligence (AI). FalconCode is provided free-of-charge, and is available upon request for computer science education research.},
booktitle = {Proceedings of the 54th ACM Technical Symposium on Computer Science Education V. 1},
pages = {938–944},
numpages = {7},
keywords = {computer science education, dataset, student code repository},
location = {Toronto ON, Canada},
series = {SIGCSE 2023}
}