Dataset Card for PyMETA

Dataset Summary

PyMETA (Python Multi-Error TAxonomy) is a large-scale benchmark for hierarchical student code error classification. It contains 48,646 real Python code submissions from 579 users on 153 question IDs (75 distinct problems; several problems were reused across course offerings under different IDs), collected from the Circle Cat online learning platform. Every submission has a single-error label derived from Online Judge execution output, organized under a three-level hierarchical taxonomy (binary → three-class → 14 fine-grained error types grounded in Python's official exception hierarchy).

An expert-annotated multi-error diagnostic subset of 97 samples (13 error types) is provided in the GitHub repository for studying co-occurring errors; metadata/multi_error_subset_97_ids.csv in this repository locates those samples by attemptstepid.

Supported Tasks

• Task A — Binary classification: No Error vs. Error.
• Task B — Three-class classification: No Error / Logic Error / Explicit Error.
• Task C — Multi-class classification: 14 fine-grained error types.
• Multi-error classification (on the 97-sample subset): identify all concurrent errors in a submission.

Languages

Student code is Python. Problem descriptions are primarily in Chinese (zh); the evaluation prompts are in English (en).

How to use

from datasets import load_dataset

ds = load_dataset("CircleCat/pymeta")                    # v2.0: question-ID split (this card)
ds_v1 = load_dataset("CircleCat/pymeta", revision="v1.0")  # v1.0: original random split

# Build the labels for Tasks A/B/C from the label columns
LABEL_IDS = {
    "No error": 0, "LogicError": 1, "SyntaxError": 2, "NameError": 3, "TypeError": 4,
    "IndentationError": 5, "UnboundLocalError": 6, "KeyError": 7, "IndexError": 8,
    "EOFError": 9, "RecursionError": 10, "ValueError": 11, "TabError": 12,
}  # any other value -> 13 (Other Errors)

def add_labels(ex):
    c = LABEL_IDS.get(ex["all_errortype2"], 13)
    return {"task_a": int(c != 0), "task_b": 0 if c == 0 else (1 if c == 1 else 2), "task_c": c}

ds = ds.map(add_labels)

When showing a problem to a model, combine question with answer_preload: many descriptions refer to "the code below" (以下代码), and that code is the starter code in answer_preload, not part of question.

Dataset Structure

Data Fields

Field	Type	Description
Unnamed: 0	int	Row index in the original full dataset (0–48,645); unique and stable across versions.
userId	int	Numeric identifier of the student.
name	string	Lesson / quiz name in the course (71 values, e.g. 循环LAB3).
questionId	int	Question identifier (153 values). Some problems appear under several IDs; see metadata/question_groups.csv.
question	string	Problem description as shown to students. Does not include the starter code (see answer_preload).
answer_preload	string	New in v2.0. Starter code pre-filled in the student's code editor when the question was opened (CodeRunner answer preload): input-reading scaffolds, fill-in-the-blank skeletons, or buggy code to fix for debugging questions. Empty (loaded as None) for the 26 questions without starter code.
exceptedAnswer	string	Reference solution (column name kept from v1.0).
attemptId	int	Question-attempt identifier; one attempt can contain several submissions (rows).
state	string	Platform state of the submission step: todo, complete, or invalid.
studentAnswer	string	The student's full submitted code. Because the editor starts with the starter code, this usually already contains answer_preload (possibly edited by the student). Uses \r\n line endings, while answer_preload uses \n.
testOutcome	string	Raw Online Judge (CodeRunner) testing outcome, PHP-serialized: per-test results, marks and error messages.
attemptstepid	int	Submission (attempt-step) identifier; unique per row across all splits and versions.
status	int	Judge status: 1 code ran, 2 compile-time error, 4 sandbox error (code was not executed).
R_errorcount	int	Number of failed test cases.
R_traceback	string	Runtime traceback (No error if none).
R_errortype	string	Runtime exception type (No error if none).
C_errormessage	string	Compile-time error message (No error if none).
C_errortype	string	Compile-time error type: SyntaxError, IndentationError, TabError, SyntaxWarning, or No error.
all_errortype	string	Exception type from either stage (C_errortype if present, otherwise R_errortype). Code that runs without an exception but fails tests is No error here.
all_errortype2	string	Source of the Task B/C labels. Same as all_errortype, except that submissions which raise no exception but fail at least one test are LogicError. The six rare types (AttributeError, RuntimeError, SyntaxWarning, ZeroDivisionError, MemoryError, ModuleNotFoundError; 178 rows) form Other Errors.
error_category	string	Task A label: error / no_error.

Taxonomy (label IDs)

0 No Error · 1 Logic Error · 2 Syntax Error · 3 Name Error · 4 Type Error · 5 Indentation Error · 6 Unbound Local Error · 7 Key Error · 8 Index Error · 9 EOF Error · 10 Recursion Error · 11 Value Error · 12 Tab Error · 13 Other Errors. Task B groups IDs 2–13 as Explicit Error. Full definitions and per-class counts are in the GitHub TAXONOMY.md.

Data Splits

Split	Examples	Question IDs	Error rate
train	38,918	133	49.1%
validation	4,863	11	61.4%
test	4,865	9	69.0%
total	48,646	153	52.3%

Since v2.0 the splits are made at the question-ID level: no questionId and no attemptId appears in more than one split. All 14 Task C classes occur in test; the validation split has no KeyError samples. Label distributions differ between splits.

Note that the course reused some problems under several question IDs (153 IDs correspond to 75 distinct problems, matched by identical description or reference solution). As a result, all 9 test question IDs and 7 of the 11 validation question IDs have a duplicate problem in train (100% of test rows, 82.7% of validation rows), and most students appear in every split. This does not affect zero-shot prompting evaluation, but matters for models trained on train. For problem-disjoint experiments, re-split using the problem_group column of metadata/question_groups.csv.

Metadata files

• metadata/question_groups.csv: questionId → problem_group (75 groups), with each ID's split, row count and lesson name.
• metadata/multi_error_subset_97_ids.csv: the 97 expert-annotated multi-error samples with their attemptstepid, questionId, attemptId, userId, gold single label, annotated labels, and v2.0 split (74 train, 6 validation, 17 test). Sample_Index is the original position in the v1.0 test split, as used by the GitHub scripts.

Known Issues

These come from the original course material or the judge and are documented rather than changed, so that every row's content and labels are identical to v1.0.

• Questions 215, 241 (train) and 519 (validation), 134 rows: question and exceptedAnswer ask for score_dict / status_dict, while answer_preload and almost all student answers print my_dict[test_param]. The problem appears to have been edited during the course; the description and reference solution reflect a later version than the one students solved.
• Question 236 (train, 74 rows): the starter code uses input().split(''), which always raises ValueError: empty separator. 23 rows labelled ValueError come from this starter-code line rather than from student-written code.
• Question 2923 (test, 17 rows): the example outputs in the description (2 for 1 2 3, 7 for 6 7 8) contradict the explanations and the reference solution, which compute the sum (6, 21).
• Sandbox errors (5 rows in train, status == 4, state == invalid): the code was not executed because the judge sandbox was unavailable, but the rows are labelled No error (attemptstepid 39969, 234690, 234936, 234945, 235723).
• Formatting artifacts in question: 9 descriptions contain hint text rendered with combining strikethrough characters (U+0336), and a few contain leftover HTML artifacts such as   or code-block line numbers.

Source Data

Submissions were collected from historical logs of the Circle Cat online learning platform (a self-hosted Moodle instance with an integrated Online Judge). They were generated organically by learners of varying proficiency during ordinary coursework, so they provide a realistic and diverse distribution of student code errors.

Personal and Sensitive Information

The data were collected from pre-existing educational records under the platform's terms of use (which inform users that anonymized data may be used for educational and research purposes), with no recruitment or experimental intervention. The dataset contains no names, e-mail addresses or other direct identifiers; students appear only as numeric userIds. The dataset contains no offensive content.

Versions

• v2.0 (current): question-ID-level re-split (made in February 2026 in response to review); added the answer_preload column and the metadata/ files; documented all columns, the label derivation and known issues. Row contents are identical to v1.0; only the split assignment changed.
• v1.0 (tag v1.0, commit 2279d02): random submission-level split stratified by error_category (39,402 / 4,379 / 4,865).

Licensing Information

Released under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license.

Citation

@inproceedings{
    tang2026pymeta,
    title={Py{META}: Evaluating Student Code Diagnosis on and Beyond the First Execution Error},
    author={Ziqi Tang and Chuyue Li and Jingyi Wang and Yu Wu and Kazuma Hashimoto and Lingyu Gao},
    booktitle={NeurIPS 2026 Trust-AI-Eval Workshop},
    year={2026},
    url={https://openreview.net/forum?id=Fz4nqncIlh}
}