XES3G5M¶
XES3G5M is a knowledge tracing benchmark dataset collected from TAL Education Xueersi 3rd-grade mathematics courses, providing question- and knowledge-concept-level sequences alongside auxiliary question texts, explanations, and images.
Dataset Link¶
Official repository: ai4ed/XES3G5M
Download portal: XES3G5M Data Package
Challenge details: AAAI2023 Global Knowledge Tracing Challenge
Data Scale and Format¶
Metrics |
Scale |
Description |
|---|---|---|
Student Count |
18,066 |
14,453 in train/validation sets, 3,613 in independent test set |
Question Count |
7,652 |
Valid questions contained in |
Knowledge Concept Path Count |
1,175 |
Standard knowledge concept paths mapped in |
Question-Level Sequence Count |
34,578 |
30,965 train/validation windows, 3,613 test set windows |
Concept-Level Sequence Count |
37,010 |
33,397 train/validation windows, 3,613 test set windows |
Note
XES3G5M originates from the AAAI 2023 Global Knowledge Tracing Challenge; the official release is tailored for sequential modeling, pre-split into windows with maximum length 200. UniKT loads and evaluates on question-level sequences question_level by default.
Release Package Fields¶
The following fields are grouped by actual table headers in each CSV of the release package.
Base Columns Shared Across All Sequence Files¶
Field Name |
Data Type |
Description |
|---|---|---|
|
Integer |
Fold index for train/val sets, values 0 to 4; -1 for test set |
|
Integer |
Student ID |
|
String |
Question ID sequence, comma-separated |
|
String |
Knowledge concept ID sequence, comma-separated; multiple concepts separated by underscores in question-level files |
|
String |
Response correctness sequence, 1 for correct, 0 for incorrect |
|
String |
Response start timestamp sequence, in milliseconds |
Additional Columns by File¶
Field Name |
Appearing Files |
Description |
|---|---|---|
|
Question-level train and window test files, |
Batch valid bit mask; -1 indicates padding or historical-only interactions excluded from loss and evaluation |
|
|
1 indicates current concept belongs to the same question as preceding concept, 0 indicates a new question |
|
|
Interaction index sequence in concept-level test set |
|
|
Interaction index sequence in question-level test set |
|
|
Remaining knowledge concepts count for this interaction in original sequence, used for prediction fusion |
|
|
Row index of this interaction in the test set |
Question-level test file question_level/test_quelevel.csv contains base columns only, excluding selectmasks, is_repeat, cidxs, qidxs, rest, orirow.
Question Metadata (metadata/questions.json)¶
Field Name |
Data Type |
Description |
|---|---|---|
|
String |
Question stem text, preserving LaTeX |
|
String |
Knowledge concept path text |
|
String |
Standard correct answer |
|
List[String] |
Multiple choice options, can be null for fill-in-the-blank questions |
|
String |
Explanation text, preserving LaTeX |
|
Integer |
Question type, 0 for fill-in-the-blank, 1 for multiple choice |
metadata/kc_routes_map.json records mappings between knowledge concept IDs and syllabus path texts; metadata/embeddings/ and metadata/images/ provide vector representations and images for questions, explanations, and concepts respectively.